1: %!PS-Adobe-3.0
2: %%Creator: groff version 1.10
3: %%CreationDate: Fri Jan 19 11:35:54 1996
4: %%DocumentNeededResources: font Times-Roman
5: %%+ font Times-Italic
6: %%+ font Times-Bold
7: %%+ font Symbol
8: %%DocumentSuppliedResources: procset grops 1.10 0
9: %%+ font Symbol-Slanted
10: %%Pages: 26
11: %%PageOrder: Ascend
12: %%Orientation: Portrait
13: %%EndComments
14: %%BeginProlog
15: %%BeginResource: procset grops 1.10 0
16: /setpacking where{
17: pop
18: currentpacking
19: true setpacking
20: }if
21: /grops 120 dict dup begin
22: /SC 32 def
23: /A/show load def
24: /B{0 SC 3 -1 roll widthshow}bind def
25: /C{0 exch ashow}bind def
26: /D{0 exch 0 SC 5 2 roll awidthshow}bind def
27: /E{0 rmoveto show}bind def
28: /F{0 rmoveto 0 SC 3 -1 roll widthshow}bind def
29: /G{0 rmoveto 0 exch ashow}bind def
30: /H{0 rmoveto 0 exch 0 SC 5 2 roll awidthshow}bind def
31: /I{0 exch rmoveto show}bind def
32: /J{0 exch rmoveto 0 SC 3 -1 roll widthshow}bind def
33: /K{0 exch rmoveto 0 exch ashow}bind def
34: /L{0 exch rmoveto 0 exch 0 SC 5 2 roll awidthshow}bind def
35: /M{rmoveto show}bind def
36: /N{rmoveto 0 SC 3 -1 roll widthshow}bind def
37: /O{rmoveto 0 exch ashow}bind def
38: /P{rmoveto 0 exch 0 SC 5 2 roll awidthshow}bind def
39: /Q{moveto show}bind def
40: /R{moveto 0 SC 3 -1 roll widthshow}bind def
41: /S{moveto 0 exch ashow}bind def
42: /T{moveto 0 exch 0 SC 5 2 roll awidthshow}bind def
43: /SF{
44: findfont exch
45: [exch dup 0 exch 0 exch neg 0 0]makefont
46: dup setfont
47: [exch/setfont cvx]cvx bind def
48: }bind def
49: /MF{
50: findfont
51: [5 2 roll
52: 0 3 1 roll
53: neg 0 0]makefont
54: dup setfont
55: [exch/setfont cvx]cvx bind def
56: }bind def
57: /level0 0 def
58: /RES 0 def
59: /PL 0 def
60: /LS 0 def
61: /MANUAL{
62: statusdict begin/manualfeed true store end
63: }bind def
64: /PLG{
65: gsave newpath clippath pathbbox grestore
66: exch pop add exch pop
67: }bind def
68: /BP{
69: /level0 save def
70: 1 setlinecap
71: 1 setlinejoin
72: 72 RES div dup scale
73: LS{
74: 90 rotate
75: }{
76: 0 PL translate
77: }ifelse
78: 1 -1 scale
79: }bind def
80: /EP{
81: level0 restore
82: showpage
83: }bind def
84: /DA{
85: newpath arcn stroke
86: }bind def
87: /SN{
88: transform
89: .25 sub exch .25 sub exch
90: round .25 add exch round .25 add exch
91: itransform
92: }bind def
93: /DL{
94: SN
95: moveto
96: SN
97: lineto stroke
98: }bind def
99: /DC{
100: newpath 0 360 arc closepath
101: }bind def
102: /TM matrix def
103: /DE{
104: TM currentmatrix pop
105: translate scale newpath 0 0 .5 0 360 arc closepath
106: TM setmatrix
107: }bind def
108: /RC/rcurveto load def
109: /RL/rlineto load def
110: /ST/stroke load def
111: /MT/moveto load def
112: /CL/closepath load def
113: /FL{
114: currentgray exch setgray fill setgray
115: }bind def
116: /BL/fill load def
117: /LW/setlinewidth load def
118: /RE{
119: findfont
120: dup maxlength 1 index/FontName known not{1 add}if dict begin
121: {
122: 1 index/FID ne{def}{pop pop}ifelse
123: }forall
124: /Encoding exch def
125: dup/FontName exch def
126: currentdict end definefont pop
127: }bind def
128: /DEFS 0 def
129: /EBEGIN{
130: moveto
131: DEFS begin
132: }bind def
133: /EEND/end load def
134: /CNT 0 def
135: /level1 0 def
136: /PBEGIN{
137: /level1 save def
138: translate
139: div 3 1 roll div exch scale
140: neg exch neg exch translate
141: 0 setgray
142: 0 setlinecap
143: 1 setlinewidth
144: 0 setlinejoin
145: 10 setmiterlimit
146: []0 setdash
147: /setstrokeadjust where{
148: pop
149: false setstrokeadjust
150: }if
151: /setoverprint where{
152: pop
153: false setoverprint
154: }if
155: newpath
156: /CNT countdictstack def
157: userdict begin
158: /showpage{}def
159: }bind def
160: /PEND{
161: clear
162: countdictstack CNT sub{end}repeat
163: level1 restore
164: }bind def
165: end def
166: /setpacking where{
167: pop
168: setpacking
169: }if
170: %%EndResource
171: %%IncludeResource: font Symbol
172: %%IncludeResource: font Times-Roman
173: %%IncludeResource: font Times-Italic
174: %%IncludeResource: font Times-Bold
175: %%BeginResource: font Symbol-Slanted
176: %%DocumentNeededResources: font Symbol
177: /MakeTransformedFont{
178: findfont dup maxlength dict begin
179: {
180: exch dup dup/FID ne exch/UniqueID ne and{
181: exch def
182: }{
183: pop pop
184: }ifelse
185: }forall
186: /FontBBox
187: currentdict/FontBBox get
188: 4 array copy def
189: FontBBox aload pop
190: 4 index transform 4 2 roll
191: 4 index transform 4 2 roll
192: FontBBox astore pop
193: FontMatrix exch matrix concatmatrix
194: /FontMatrix exch def
195: dup/FontName exch def
196: currentdict end
197: definefont pop
198: }bind def
199: /Symbol-Slanted
200: [.89 0.0 15.5 dup sin exch cos div .89 0.0 0.0]
201: /Symbol
202: MakeTransformedFont
203: %%EndResource
204: grops begin/DEFS 1 dict def DEFS begin/u{.001 mul}bind def end/RES 72
205: def/PL 792 def/LS false def/ENC0[/asciicircum/asciitilde/Scaron/Zcaron
206: /scaron/zcaron/Ydieresis/trademark/quotesingle/.notdef/.notdef/.notdef
207: /.notdef/.notdef/.notdef/.notdef/.notdef/.notdef/.notdef/.notdef/.notdef
208: /.notdef/.notdef/.notdef/.notdef/.notdef/.notdef/.notdef/.notdef/.notdef
209: /.notdef/.notdef/space/exclam/quotedbl/numbersign/dollar/percent
210: /ampersand/quoteright/parenleft/parenright/asterisk/plus/comma/hyphen
211: /period/slash/zero/one/two/three/four/five/six/seven/eight/nine/colon
212: /semicolon/less/equal/greater/question/at/A/B/C/D/E/F/G/H/I/J/K/L/M/N/O
213: /P/Q/R/S/T/U/V/W/X/Y/Z/bracketleft/backslash/bracketright/circumflex
214: /underscore/quoteleft/a/b/c/d/e/f/g/h/i/j/k/l/m/n/o/p/q/r/s/t/u/v/w/x/y
215: /z/braceleft/bar/braceright/tilde/.notdef/quotesinglbase/guillemotleft
216: /guillemotright/bullet/florin/fraction/perthousand/dagger/daggerdbl
217: /endash/emdash/ff/fi/fl/ffi/ffl/dotlessi/dotlessj/grave/hungarumlaut
218: /dotaccent/breve/caron/ring/ogonek/quotedblleft/quotedblright/oe/lslash
219: /quotedblbase/OE/Lslash/.notdef/exclamdown/cent/sterling/currency/yen
220: /brokenbar/section/dieresis/copyright/ordfeminine/guilsinglleft
221: /logicalnot/minus/registered/macron/degree/plusminus/twosuperior
222: /threesuperior/acute/mu/paragraph/periodcentered/cedilla/onesuperior
223: /ordmasculine/guilsinglright/onequarter/onehalf/threequarters
224: /questiondown/Agrave/Aacute/Acircumflex/Atilde/Adieresis/Aring/AE
225: /Ccedilla/Egrave/Eacute/Ecircumflex/Edieresis/Igrave/Iacute/Icircumflex
226: /Idieresis/Eth/Ntilde/Ograve/Oacute/Ocircumflex/Otilde/Odieresis
227: /multiply/Oslash/Ugrave/Uacute/Ucircumflex/Udieresis/Yacute/Thorn
228: /germandbls/agrave/aacute/acircumflex/atilde/adieresis/aring/ae/ccedilla
229: /egrave/eacute/ecircumflex/edieresis/igrave/iacute/icircumflex/idieresis
230: /eth/ntilde/ograve/oacute/ocircumflex/otilde/odieresis/divide/oslash
231: /ugrave/uacute/ucircumflex/udieresis/yacute/thorn/ydieresis]def
232: /Times-Bold@0 ENC0/Times-Bold RE/Times-Italic@0 ENC0/Times-Italic RE
233: /Times-Roman@0 ENC0/Times-Roman RE
234: %%EndProlog
235: %%Page: 1 1
236: %%BeginPageSetup
237: gsave %matrix defaultmatrix setmatrix
238: 90 rotate 4 72 mul .55 -72 mul moveto /Times-Roman findfont
239: 20 scalefont setfont 0.3 setgray (arXiv:cs.MS/0001018 23 Jan 2000) show grestore
240: BP
241: %%EndPageSetup
242: /F0 11/Times-Roman@0 SF(%A L. Ingber)72 85 Q(%T Adapti)72 98 Q .33 -.165
243: (ve s)-.275 H(imulated annealing \(ASA\): Lessons learned).165 E
244: (%J Control and Cybernetics)72 111 Q(%D 1995)72 124 Q
245: (%P \(to be published\))72 137 Q 1.039(This is an in)72 150 R 1.038
246: (vited paper to a special issue of the Polish Journal)-.44 F/F1 11
247: /Times-Italic@0 SF(Contr)3.788 E 1.038(ol and Cybernetics)-.495 F F0
248: 1.038(on \231Simulated)3.788 F
249: (Annealing Applied to Combinatorial Optimization.)72 163 Q<9a>-.77 E/F2
250: 11/Times-Bold@0 SF(Adapti)179.533 192.9 Q .22 -.11(ve s)-.11 H
251: (imulated annealing \(ASA\): Lessons lear).11 E(ned)-.165 E F0
252: (Lester Ingber)276.525 209.8 Q(Lester Ingber Research)254.993 235.8 Q
253: -1.221(P.)282.163 248.8 S(O.B. 857)1.221 E(McLean, V)262.599 261.8 Q
254: 2.75(A2)-1.485 G(2101)-2.75 E(U.S.A.)290.875 274.8 Q
255: (ingber@alumni.caltech.edu)245.335 287.8 Q F2(ABSTRA)99.5 331.8 Q(CT)
256: -.605 E F0(Adapti)127 357.7 Q 1.779 -.165(ve s)-.275 H 1.449(imulated a\
257: nnealing \(ASA\) is a global optimization algorithm based on an).165 F
258: 2.018
259: (associated proof that the parameter space can be sampled much more ef)
260: 99.5 379.7 R 2.017(\214ciently than by)-.275 F 2.713(using other pre)
261: 99.5 401.7 R 2.713(vious simulated annealing algorithms.)-.275 F 2.714
262: (The author')8.213 F 5.464(sA)-.605 G 2.714(SA code has been)-5.464 F
263: 1.432(publicly a)99.5 423.7 R -.275(va)-.22 G 1.432(ilable for o).275 F
264: -.165(ve)-.165 G 4.182(rt).165 G 1.652 -.11(wo y)-4.182 H 4.182
265: (ears. During).11 F 1.432(this time the author has v)4.182 F 1.432
266: (olunteered to help)-.22 F 2.038(people via e-mail, and the feedback ob\
267: tained has been used to further de)99.5 445.7 R -.165(ve)-.275 G 2.039
268: (lop the code.).165 F 1.85
269: (Some lessons learned, in particular some which are rele)99.5 467.7 R
270: -.275(va)-.275 G 1.85(nt to other simulated annealing).275 F
271: (algorithms, are described.)99.5 489.7 Q EP
272: %%Page: 2 2
273: %%BeginPageSetup
274: BP
275: %%EndPageSetup
276: /F0 11/Times-Roman@0 SF(ASA Lessons Learned)72 58 Q 165.888(-2- Lester)
277: 128.609 F(Ingber)2.75 E/F1 11/Times-Bold@0 SF 2.75(1. INTR)72 94 R
278: (ODUCTION)-.33 E F0(Adapti)72 119.9 Q 4.383 -.165(ve s)-.275 H 4.053(im\
279: ulated annealing \(ASA\) is a global optimization algorithm that relies\
280: on randomly).165 F .201(importance-sampling the parameter space, i.e.,\
281: in contrast to utilizing deterministic approaches often used)72 141.9 R
282: (by OR and mathematical programming people.)72 163.9 Q 1.204
283: (Since the public release of the v)99.5 189.8 R 1.204(ery f)-.165 F
284: 1.204(ast simulated reannealing \(VFSR\) code \(Ingber)-.11 F 3.954(,1)
285: -.44 G 1.205(989\), no)-3.954 F(w)-.275 E .956(called ASA \(Ingber)72
286: 211.8 R 3.706(,1)-.44 G .956(993a\), I ha)-3.706 F 1.286 -.165(ve v)-.22
287: H .956
288: (olunteered to help users via e-mail on code-speci\214c problems the)
289: -.055 F(y)-.165 E 1.352(might encounter)72 233.8 R 6.852(.T)-.605 G
290: 1.353(he popularity of the code, roughly measured by the size of the AS\
291: A_list of people)-6.852 F 1.282(speci\214cally requesting to be placed \
292: on a monthly list of updates\212no)72 255.8 R 4.032(wa)-.275 G 4.032(ta)
293: -4.032 G 1.281(bout 500 names\212increased)-4.032 F .631(about a f)72
294: 277.8 R .631(actor of tw)-.11 F 3.381(oa)-.11 G .631(fter an article in)
295: -3.381 F/F2 11/Times-Italic@0 SF .631(The W)3.381 F .631(all Str)-1.012
296: F .631(eet J)-.407 F(ournal)-.275 E F0(\(W)3.381 E(ofse)-.88 E 2.062
297: -.715(y, 1)-.165 H 3.382(993\). I).715 F(ha)3.382 E .962 -.165(ve h)-.22
298: H .632(ad contact with).165 F .03(thousands of users via e-mail, a fe)72
299: 299.8 R 2.779(wv)-.275 G .029(ia other channels of communication, and i\
300: t is clear that these are just)-2.779 F 3.548(as)72 321.8 S .798
301: (mall fraction of people that ha)-3.548 F 1.128 -.165(ve a)-.22 H 3.548
302: (tl).165 G .798(east tried to use the code.)-3.548 F .799
303: (Therefore, I conclude that the code is)6.298 F
304: (quite stable and free of major b)72 343.8 Q(ugs.)-.22 E .099(This pape\
305: r deals with some of the lessons learned from this interaction.)99.5
306: 369.7 R 2.848(It)5.599 G .098(hink that at least some of)-2.848 F 1.327
307: (these lessons may be useful to other de)72 391.7 R -.165(ve)-.275 G
308: 1.328(lopers of simulated annealing \(SA\) code, as well as to man).165
309: F(y)-.165 E 4.689(users. A)72 413.7 R(pre)4.689 E 1.938
310: (vious paper has described SA in the conte)-.275 F 1.938
311: (xt of its use across man)-.165 F 4.688(yd)-.165 G 1.938
312: (isciplines, and the)-4.688 F -.275(va)72 435.7 S
313: (riation in use in actual practice v).275 E(ersus the algorithm of)-.165
314: E(fered in theoretical papers \(Ingber)-.275 E 2.75(,1)-.44 G(993b\).)
315: -2.75 E .812(Section 2 gi)99.5 461.6 R -.165(ve)-.275 G -6.311 3.562
316: (sa b).165 H .813(rief mathematical description of the ASA algorithm.)
317: -3.562 F .813(Section 3 describes some)6.313 F .595
318: (of the options in the code, and e)72 483.6 R .595(xplains wh)-.165 F
319: 3.345(yt)-.055 G(he)-3.345 E 3.345(ya)-.165 G .595(re useful.)-3.345 F
320: .595(Section 4 addresses some comparison tests)6.095 F .886(and contro)
321: 72 505.6 R -.165(ve)-.165 G .886(rsial claims, emphasizing that in f)
322: .165 F .887(act there are man)-.11 F 3.637(ys)-.165 G .887(hades of SA.)
323: -3.637 F .887(Section 5 gi)6.387 F -.165(ve)-.275 G 3.637(sab).165 G
324: (rief)-3.637 E(conclusion.)72 527.6 Q .377
325: (ASA source code in C-language is publicly a)99.5 553.5 R -.275(va)-.22
326: G .377(ilable via anon).275 F .376(ymous ftp \(Ingber)-.165 F 3.126(,1)
327: -.44 G .376(993a\), or via the)-3.126 F -.11(wo)72 575.5 S 2.405
328: (rld wide web \(WWW\) at http://www).11 F 5.155
329: (.alumni.caltech.edu/~ingber/. Problems)-.715 F 2.406
330: (with the code can be)5.156 F 2.53
331: (addressed to ingber@alumni.caltech.edu.)72 597.5 R 2.529
332: (Requests to be placed on the ASA mailing list should be)8.03 F
333: (addressed to asa\255request@alumni.caltech.edu.)72 619.5 Q F1 2.75
334: (2. ASA)72 663.5 R(ALGORITHM)2.75 E EP
335: %%Page: 3 3
336: %%BeginPageSetup
337: BP
338: %%EndPageSetup
339: /F0 11/Times-Roman@0 SF(ASA Lessons Learned)72 58 Q 165.888(-3- Lester)
340: 128.609 F(Ingber)2.75 E/F1 11/Times-Bold@0 SF 2.75
341: (2.1. \231Standard\232)72 94 R(Simulated Annealing \(SA\))2.75 E F0 .533
342: (The Metropolis Monte Carlo inte)99.5 119.9 R .533
343: (gration algorithm \(Metropolis)-.165 F/F2 11/Times-Italic@0 SF .534
344: (et al)3.283 F F0 3.284(,1)C .534(953\) w)-3.284 F .534
345: (as generalized by the)-.11 F .702
346: (Kirkpatrick algorithm to include a temperature schedule for ef)72 141.9
347: R .702(\214cient searching \(Kirkpatrick)-.275 F F2 .701(et al)3.452 F
348: F0 3.451(,1)C(983\).)-3.451 E 4.028(As)72 163.9 S(uf)-4.028 E(\214cienc)
349: -.275 E 4.028(yp)-.165 G 1.278(roof w)-4.028 F 1.278(as then sho)-.11 F
350: 1.278(wn to put an lo)-.275 F 1.279(wer bound on that schedule as 1/)
351: -.275 F(log\()1.87 E F2(t).143 E F0 1.279(\), where).748 F F2(t)4.172 E
352: F0 1.279(is an)4.777 F 2.133(arti\214cial time measure of the annealing\
353: schedule \(Geman and Geman, 1984\).)72 185.9 R(Ho)7.632 E(we)-.275 E
354: -.165(ve)-.275 G 3.012 -.44(r, i).165 H(ndependent).44 E 2.566
355: (credit usually goes to se)72 207.9 R -.165(ve)-.275 G 2.566
356: (ral other authors for independently de).165 F -.165(ve)-.275 G 2.567
357: (loping the algorithm that is no).165 F(w)-.275 E
358: (recognized as simulated annealing \(Cern)72 229.9 Q 1.43 -.715(y, 1)
359: -.165 H(982; Pincus, 1970\).).715 E F1 2.75(2.1.1. Boltzmann)72 273.9 R
360: (annealing \(B)2.75 E(A\))-.33 E F0 1.228
361: (Credit for the \214rst simulated annealing is generally gi)99.5 303.7 R
362: -.165(ve)-.275 G 3.977(nt).165 G 3.977(oaM)-3.977 G 1.227
363: (onte Carlo importance-sampling)-3.977 F .385(technique for doing lar)72
364: 325.7 R .385(ge-dimensional path inte)-.198 F .385
365: (grals arising in statistical ph)-.165 F .386
366: (ysics problems \(Metropolis)-.055 F F2(et)3.136 E(al)72 347.7 Q F0
367: 5.295(,1)C 5.295(953\). This)-5.295 F 2.545(method w)5.295 F 2.545
368: (as generalized to \214tting non-con)-.11 F 2.874 -.165(vex c)-.44 H
369: 2.544(ost-functions arising in a v).165 F 2.544(ariety of)-.275 F 1.239
370: (problems, e.g., \214nding the optimal wiring for a densely wired compu\
371: ter chip \(Kirkpatrick)72 369.7 R F2 1.239(et al)3.989 F F0 3.989(,1)C
372: (983\).)-3.989 E 1.635(The choices of probability distrib)72 391.7 R
373: 1.634(utions described in this section are generally speci\214ed as Bol\
374: tzmann)-.22 F(annealing \(B)72 413.7 Q(A\) \(Szu and Hartle)-.385 E 1.43
375: -.715(y, 1)-.165 H(987\).).715 E(The method of simulated annealing cons\
376: ists of three functional relationships.)99.5 439.6 Q(1.)99.5 462.102 Q
377: F2(g)5.202 E/F3 8/Times-Italic@0 SF(T).17 2.2 M F0(\()1.566 -2.2 M F2(x)
378: .847 E F0 1.99(\): Probability density of state-space of).583 F F2(D)
379: 5.378 E F0(parameters)5.048 E F2(x)5.587 E/F4 11/Symbol SF(=)3.663 E F2
380: .847({x)3.08 G F3(i)-.256 -4.62 M F0(;).838 4.62 M F2(i)1.881 E F4(=)
381: 3.476 E F0(1,)3.08 E F2 .308(D})2.508 G F0 4.741(,w)-.308 G 1.991
382: (here the)-4.741 F(subscript)99.5 484.102 Q F2(T)2.651 E F0(signi\214es\
383: a parameterization popularly referred to as the \231temperature.)4.147
384: E<9a>-.77 E(2.)99.5 506.102 Q F2(h)3.091 E F0(\().308 E F4(D)A F2(E).561
385: E F0(\): Probability for acceptance of ne).803 E 2.75(wc)-.275 G
386: (ost-function gi)-2.75 E -.165(ve)-.275 G 2.75(nt).165 G(he just pre)
387: -2.75 E(vious v)-.275 E(alue.)-.275 E(3.)99.5 528.102 Q F2(T)5.049 E F0
388: (\()1.397 E F2(k).396 E F0 2.398
389: (\): schedule of \231annealing\232 the \231temperature\232).737 F F2(T)
390: 5.049 E F0 2.397(in annealing-time steps)6.544 F F2(k)5.543 E F0 5.147
391: (,i).737 G 2.397(.e., of)-5.147 F .058(changing the v)99.5 550.102 R
392: .058(olatility or \215uctuations of one or both of the tw)-.22 F 2.809
393: (op)-.11 G(re)-2.809 E .059(vious probability densities.)-.275 F 1.335
394: (The acceptance probability is based on the chances of obtaining a ne)
395: 99.5 576.002 R 4.085(ws)-.275 G 1.335(tate with \231ener)-4.085 F
396: (gy\232)-.198 E F2(E)4.646 E F3(k).761 2.2 M/F5 8/Symbol SF(+).536 E/F6
397: 8/Times-Roman@0 SF(1)A F0(relati)72 598.002 Q .33 -.165(ve t)-.275 H
398: 2.75(oap).165 G(re)-2.75 E(vious state with \231ener)-.275 E(gy\232)
399: -.198 E F2(E)3.311 E F3(k).761 2.2 M F0(,)1.086 -2.2 M F2(h)108.341
400: 632.288 Q F0(\().308 E F4(D)A F2(E).561 E F0(\)).803 E F4(=)3.08 E F0
401: -.165(ex)35.958 -7.7 O(p\().165 E F4(-)A F2(E).561 E F3(k).761 2.2 M F5
402: (+).536 E F6(1)A F0(/).55 -2.2 M F2(T)-.099 E F0(\))1.397 E -.165(ex)
403: -92.187 15.4 O(p\().165 E F4(-)A F2(E).561 E F3(k).761 2.2 M F5(+).536 E
404: F6(1)A F0(/).55 -2.2 M F2(T)-.099 E F0(\))1.397 E F4(+)2.42 E F0 -.165
405: (ex)2.42 G(p\().165 E F4(-)A F2(E).561 E F3(k).761 2.2 M F0(/)1.086 -2.2
406: M F2(T)-.099 E .44 LW 273.556 629.428 149.811 629.428 DL F0(\))269.893
407: 639.988 Q F4(=)139.372 679.311 Q F0(1)34.006 -7.7 M(1)-35.106 15.4 M F4
408: (+)2.42 E F0 -.165(ex)2.42 G(p\().165 E F4(D)A F2(E).561 E F0(/).803 E
409: F2(T)-.099 E 214.524 676.451 149.811 676.451 DL F0(\))210.861 687.011 Q
410: F4<bb>139.372 720.548 Q F0 -.165(ex)3.08 G(p\().165 E F4(-D)A F2(E).561
411: E F0(/).803 E F2(T)-.099 E F0 309.57 3.08(\), \()1.397 H(1\))-3.08 E EP
412: %%Page: 4 4
413: %%BeginPageSetup
414: BP
415: %%EndPageSetup
416: /F0 11/Times-Roman@0 SF(ASA Lessons Learned)72 58 Q 165.888(-4- Lester)
417: 128.609 F(Ingber)2.75 E(where)72 94 Q/F1 11/Symbol SF(D)4.444 E/F2 11
418: /Times-Italic@0 SF(E).561 E F0 1.694(represents the \231ener)5.247 F
419: 1.694(gy\232 dif)-.198 F 1.694(ference between the present and pre)-.275
420: F 1.695(vious v)-.275 F 1.695(alues of the ener)-.275 F(gies)-.198 E
421: 2.339(\(considered here as cost functions\) appropriate to the ph)72 116
422: R 2.339(ysical problem, i.e.,)-.055 F F1(D)5.089 E F2(E).561 E F0(=)
423: 5.892 E F2(E)5.65 E/F3 8/Times-Italic@0 SF(k).761 2.2 M/F4 8/Symbol SF
424: (+).536 E/F5 8/Times-Roman@0 SF(1)A F1(-)2.97 -2.2 M F2(E)2.981 E F3(k)
425: .761 2.2 M F0 7.839(.T)1.086 -2.2 O(his)-7.839 E .19
426: (essentially is the Boltzmann distrib)72 138 R .191(ution contrib)-.22 F
427: .191(uting to the statistical mechanical partition function of the)-.22
428: F(system \(Binder and Stauf)72 160 Q(fer)-.275 E 2.75(,1)-.44 G(985\).)
429: -2.75 E 1.264(This sampling algorithm also can be described by consider\
430: ing: a set of states labeled by)99.5 185.9 R F2(x)4.861 E F0 4.013(,e)
431: .583 G(ach)-4.013 E 1.463(with ener)72 207.9 R(gy)-.198 E F2(e)4.422 E
432: F0(\().198 E F2(x).847 E F0 1.463(\); a set of probability distrib).583
433: F(utions)-.22 E F2(p)5.588 E F0(\().209 E F2(x).847 E F0 1.463
434: (\); and the ener).583 F 1.464(gy distrib)-.198 F 1.464(ution per state)
435: -.22 F F2(d)4.599 E F0 -3.223(\(\().847 G F2(e)3.432 E F0(\().198 E F2
436: (x).847 E F0(\)\)).583 E(\),)-3.223 E(gi)72 229.9 Q(ving an aggre)-.275
437: E -.055(ga)-.165 G(te ener).055 E(gy)-.198 E F2(E)3.311 E F0(,).803 E F3
438: (x)111.056 267.524 Q/F6 16/Symbol SF(S)-6.608 -6.6 M F2(p)3.245 -2.524 M
439: F0(\().209 E F2(x).847 E F0(\)).583 E F2(d).385 E F0 -3.223(\(\().847 G
440: F2(e)3.432 E F0(\().198 E F2(x).847 E F0(\)\)).583 E(\))-3.223 E F1(=)
441: 3.08 E F2(E)3.641 E F0 327.125(.\()3.883 G(2\))-327.125 E
442: (The principle of maximizing the entrop)72 293.362 Q -.715(y,)-.11 G F2
443: (S)3.828 E F0(,).638 E F2(S)108.363 321.862 Q F1 3.08(=-)3.718 G F3(x)
444: 1.846 9.124 M F6(S)-6.608 -6.6 M F2(p)3.245 -2.524 M F0(\().209 E F2(x)
445: .847 E F0 1.87(\)l).583 G(n[)-1.87 E F2(p)1.375 E F0(\().209 E F2(x).847
446: E F0(\)/).583 E F2(p)1.375 E .44 LW 220.885 315.372 217.475 315.372 DL
447: F0(\()211.632 321.862 Q F2(x).847 E F0 .33(\)] ,).583 F(\(3\))292.409 E
448: 107.365 350.334 103.955 350.334 DL(where)72 356.824 Q F2(x)3.749 E F0
449: .152(represents a reference state, using Lagrange multipliers \(Mathe)
450: 3.485 F .152(ws and W)-.275 F(alk)-.88 E(er)-.11 E 2.902(,1)-.44 G .152
451: (970\) to constrain)-2.902 F(the ener)72 378.824 Q(gy to a)-.198 E -.165
452: (ve)-.22 G(rage v).165 E(alue)-.275 E F2(T)2.651 E F0 2.75(,l)1.397 G
453: (eads to the most lik)-2.75 E(ely Gibbs distrib)-.11 E(ution)-.22 E F2
454: (G)2.728 E F0(\().55 E F2(x).847 E F0(\),).583 E F2(G)107.978 413.11 Q
455: F0(\().55 E F2(x).847 E F0(\)).583 E F1(=)3.08 E F0(1)5.566 -7.7 M
456: 151.461 410.25 143.629 410.25 DL F2(Z)144.245 420.81 Q F0 -.165(ex)4.29
457: -7.7 O(p\().165 E(\()-3.223 E F1(-)A F2(H).638 E F0(\()1.045 E F2(x).847
458: E F0(\)/).583 E F2(T)-.099 E F0 6.776 -3.223(\)\) ,)1.397 H(\(4\))
459: 306.215 E 2.716(in terms of the normalizing partition function)72 445.46
460: R F2(Z)6.083 E F0 5.467(,a)1.1 G 2.717(nd the Hamiltonian)-5.467 F F2(H)
461: 6.105 E F0 2.717(operator as the \231ener)6.512 F(gy\232)-.198 E
462: (function,)72 467.46 Q F2(Z)108.616 495.96 Q F1(=)4.18 E F3(x)6.136
463: 9.124 M F6(S)-6.608 -6.6 M F0 -.165(ex)1.87 -2.524 O(p\().165 E(\()
464: -3.223 E F1(-)A F2(H).638 E F0(\()1.045 E F2(x).847 E F0(\)/).583 E F2
465: (T)-.099 E F0 6.776 -3.223(\)\) .)1.397 H(\(5\))321.493 E -.165(Fo)99.5
466: 534.822 S 2.826(rs).165 G .076(uch distrib)-2.826 F .076(utions of stat\
467: es and acceptance probabilities de\214ned by functions such as)-.22 F F2
468: (h)3.166 E F0(\().308 E F1(D)A F2(E).561 E F0 .075(\), the).803 F 1.045
469: (equilibrium principle of detailed balance holds.)72 556.822 R 1.045
470: (I.e., the distrib)6.545 F 1.046(utions of states before,)-.22 F F2(G)
471: 3.774 E F0(\().55 E F2(x).847 E F3(k).761 2.2 M F0 1.046(\), and after)
472: 1.086 -2.2 N(,)-.44 E F2(G)71.978 578.822 Q F0(\().55 E F2(x).847 E F3
473: (k).761 2.2 M F4(+).536 E F5(1)A F0
474: (\), applying the acceptance criteria,).55 -2.2 M F2(h)3.091 E F0(\()
475: .308 E F1(D)A F2(E).561 E F0(\)).803 E F1(=)3.08 E F2(h)3.421 E F0(\()
476: .308 E F2(E).561 E F3(k).761 2.2 M F4(+).536 E F5(1)A F1(-)2.97 -2.2 M
477: F2(E)2.981 E F3(k).761 2.2 M F0 2.75(\)a)1.086 -2.2 O(re the same:)-2.75
478: E F2(G)107.978 607.322 Q F0(\().55 E F2(x).847 E F3(k).761 2.2 M F0(\))
479: 1.086 -2.2 M F2(h).341 E F0 -3.223(\(\().308 G F1(D)3.223 E F2(E).561 E
480: F0(\().803 E F2(x).847 E F0(\)\)).583 E(\))-3.223 E F1(=)3.08 E F2(G)
481: 3.058 E F0(\().55 E F2(x).847 E F3(k).761 2.2 M F4(+).536 E F5(1)A F0
482: 287.179 3.08(\). \().55 -2.2 P(6\))-3.08 E 2.333(This is suf)72 635.822
483: R 2.333(\214cient to establish that all states of the system can be sam\
484: pled, in theory)-.275 F 7.833(.H)-.715 G -.275(ow)-7.833 G -2.365 -.275
485: (ev e).275 H 3.212 -.44(r, t).275 H(he).44 E .107
486: (annealing schedule interrupts equilibrium e)72 657.822 R -.165(ve)-.275
487: G .108(ry time the temperature is changed, and so, at best, this must)
488: .165 F(be done carefully and gradually)72 679.822 Q(.)-.715 E .44(An im\
489: portant aspect of the SA algorithm is to pick the ranges of the paramet\
490: ers to be searched.)99.5 705.722 R(In)5.939 E 2.947(practice, computati\
491: on of continuous systems requires some discretization, so without loss \
492: of much)72 727.722 R EP
493: %%Page: 5 5
494: %%BeginPageSetup
495: BP
496: %%EndPageSetup
497: /F0 11/Times-Roman@0 SF(ASA Lessons Learned)72 58 Q 165.888(-5- Lester)
498: 128.609 F(Ingber)2.75 E 3.079(generality for applications described her\
499: e, the space can be assumed to be discretized.)72 94 R 3.078(There are)
500: 8.578 F 1.866(additional constraints that are required when dealing wit\
501: h generating and cost functions with inte)72 116 R(gral)-.165 E -.275
502: (va)72 138 S 5.237(lues. Man).275 F 5.237(yp)-.165 G 2.487
503: (ractitioners use no)-5.237 F -.165(ve)-.165 G 5.237(lt).165 G 2.486
504: (echniques to narro)-5.237 F 5.236(wt)-.275 G 2.486
505: (he range as the search progresses.)-5.236 F -.165(Fo)7.986 G(r).165 E
506: -.165(ex)72 160 S 3.263(ample, based on functional forms deri).165 F
507: -.165(ve)-.275 G 6.013(df).165 G 3.263(or man)-6.013 F 6.013(yp)-.165 G
508: -.055(hy)-6.013 G 3.264(sical systems belonging to the class of).055 F
509: (Gaussian-Mark)72 182 Q -.165(ov)-.11 G
510: (ian systems, one could choose an algorithm for).165 E/F1 11
511: /Times-Italic@0 SF(g)3.212 E F0(,).242 E F1(g)108.462 211.178 Q F0(\()
512: .242 E/F2 11/Symbol SF(D)A F1(x).847 E F0(\)).583 E F2(=)3.08 E F0(\(2)
513: 3.08 E/F3 11/Symbol-Slanted SF(p)-.088 E F1(T)1.661 E F0(\))1.397 E/F4 8
514: /Symbol SF(-)-4.62 I/F5 8/Times-Italic@0 SF(D).464 E/F6 8/Times-Roman@0
515: SF(/2).224 E F0 -.165(ex)2.42 4.62 O(p[).165 E F2(-D)A F1(x).847 E F6(2)
516: .583 -4.62 M F0(/\(2).55 4.62 M F1(T)-.099 E F0 .33(\)] ,)1.397 F(\(7\))
517: 257.8 E(where)72 239.678 Q F2(D)3.78 E F1(x).847 E F0(=)4.363 E F1(x)
518: 4.627 E F2(-)3.003 E F1(x)3.267 E F6(0).473 2.2 M F0 1.029(is the de)
519: 4.33 -2.2 N 1.029(viation of)-.275 F F1(x)4.626 E F0(from)4.362 E F1(x)
520: 4.626 E F6(0).473 2.2 M F0 1.029(\(usually tak)4.329 -2.2 N 1.029
521: (en to be the just-pre)-.11 F 1.029(viously chosen point\),)-.275 F .907
522: (proportional to a \231momentum\232 v)72 261.678 R .907
523: (ariable, and where)-.275 F F1(T)3.558 E F0 .907
524: (is a measure of the \215uctuations of the Boltzmann)5.054 F(distrib)72
525: 283.678 Q(ution)-.22 E F1(g)3.397 E F0 .185(in the)3.177 F F1(D)3.573 E
526: F0(-dimensional).308 E F1(x)3.782 E F0 2.935(-space. Gi).583 F -.165(ve)
527: -.275 G(n).165 E F1(g)3.396 E F0(\().242 E F2(D)A F1(x).847 E F0 .184
528: (\), it has been pro).583 F -.165(ve)-.165 G 2.934(n\().165 G .184
529: (Geman and Geman, 1984\))-2.934 F(that it suf)72 305.678 Q
530: (\214ces to obtain a global minimum of)-.275 E F1(E)3.311 E F0(\().803 E
531: F1(x).847 E F0 2.75(\)i).583 G(f)-2.75 E F1(T)2.651 E F0
532: (is selected to be not f)4.147 E(aster than)-.11 E F1(T)107.901 339.711
533: Q F0(\()1.397 E F1(k).396 E F0(\)).737 E F2(=)3.08 E F1(T)7.003 -7.7 M
534: F6(0).473 2.2 M F0(ln)-13.192 13.2 M .44 LW 158.721 336.851 142.276
535: 336.851 DL F1(k)153.1 347.411 Q F0 361.303(,\()5.137 -7.7 O(8\))-361.303
536: E(with)72 372.182 Q F1(T)2.651 E F6(0).473 2.2 M F0(\231lar)3.3 -2.2 M
537: (ge enough.)-.198 E<9a>-.77 E -.165(Fo)99.5 398.082 S 2.901(rt).165 G
538: .151(he purposes of this paper)-2.901 F 2.902(,ah)-.44 G .152
539: (euristic demonstration follo)-2.902 F .152(ws, to sho)-.275 F 2.902(wt)
540: -.275 G .152(hat Eq. \(8\) will suf)-2.902 F .152(\214ce to)-.275 F(gi)
541: 72 420.082 Q 1.553 -.165(ve a g)-.275 H 1.223(lobal minimum of).165 F F1
542: (E)4.534 E F0(\().803 E F1(x).847 E F0 3.973(\)\().583 G 1.223
543: (Szu and Hartle)-3.973 F 2.653 -.715(y, 1)-.165 H 3.973(987\). In).715 F
544: 1.222(order to statistically assure, i.e., requiring)3.973 F(man)72
545: 442.082 Q 3.41(yt)-.165 G .66(rials, that an)-3.41 F 3.41(yp)-.165 G .66
546: (oint in)-3.41 F F1(x)4.257 E F0 .66
547: (-space can be sampled in\214nitely often in annealing-time \(IO).583 F
548: .661(T\), it suf)-.44 F(\214ces)-.275 E 2.859(to pro)72 464.082 R 3.189
549: -.165(ve t)-.165 H 2.859
550: (hat the products of probabilities of not generating a state).165 F F1
551: (x)6.456 E F0(IO)6.192 E 5.609(Tf)-.44 G 2.859(or all annealing-times)
552: -5.609 F(successi)72 486.082 Q .33 -.165(ve t)-.275 H(o).165 E F1(k)
553: 3.146 E F6(0).473 2.2 M F0(yield zero,)3.3 -2.2 M F5(k)112.244 528.402 Q
554: /F7 16/Symbol SF(P)-7.796 -7.334 M F0(\(1)1.87 -2.524 M F2(-)2.42 E F1
555: (g)2.882 E F5(k).53 2.2 M F0(\))1.086 -2.2 M F5(k)3.368 -6.748 M F4<ae>
556: .536 E/F8 10.4/Symbol SF<a5>1.248 I F2(=)-12.863 5.5 M F0 324.094 3.08
557: (0. \()9.904 H(9\))-3.08 E(This is equi)72 554.24 Q -.275(va)-.275 G
558: (lent to).275 E F5(k)110.836 596.56 Q F7(S)-6.388 -7.334 M F1(g)2.332
559: -2.524 M F5(k).53 2.2 M(k)4.454 -8.948 M F4<ae>.536 E F8<a5>1.248 I F2
560: (=)-12.863 5.5 M/F9 14.3/Symbol SF<a5>9.904 1.716 M F0 349.329(,\()3.08
561: -1.716 O(10\))-349.329 E 1.203
562: (as seen by taking the log of Eq. \(9\) and T)72 622.398 R 1.203
563: (aylor e)-.88 F 1.204(xpanding in)-.165 F F1(g)4.416 E F5(k).53 2.2 M F0
564: 6.704(.T)1.086 -2.2 O 1.204(he problem then reduces to \214nding)-6.704
565: F F1(T)71.901 644.398 Q F0(\()1.397 E F1(k).396 E F0 4.626(\)t).737 G
566: 4.626(os)-4.626 G 1.876(atisfy Eq. \(10\).)-4.626 F 1.876(Note that, gi)
567: 7.376 F -.165(ve)-.275 G 4.626(nav).165 G 1.875(ery lar)-4.791 F 1.875
568: (ge space to sample, often at best only a \231weak\232)-.198 F(er)72
569: 666.398 Q 2.059(godicity can be assumed for this proof, and an)-.198 F
570: 4.81(ys)-.165 G 2.06(uch er)-4.81 F 2.06(godicity e)-.198 F -.165(ve)
571: -.275 G 4.81(nf).165 G 2.06(or well-understood ph)-4.81 F(ysical)-.055 E
572: (systems is an open area of research \(Ma, 1985\).)72 688.398 Q -.165
573: (Fo)99.5 714.298 S 2.75(rB).165 G(A, if)-3.135 E F1(T)2.651 E F0(\()
574: 1.397 E F1(k).396 E F0 2.75(\)i).737 G 2.75(ss)-2.75 G
575: (elected to be Eq. \(8\), then Eq. \(7\) gi)-2.75 E -.165(ve)-.275 G(s)
576: .165 E EP
577: %%Page: 6 6
578: %%BeginPageSetup
579: BP
580: %%EndPageSetup
581: /F0 11/Times-Roman@0 SF(ASA Lessons Learned)72 58 Q 165.888(-6- Lester)
582: 128.609 F(Ingber)2.75 E/F1 10.4/Symbol SF<a5>112.469 89.952 Q/F2 8
583: /Times-Italic@0 SF(k)-11.596 19.312 M/F3 8/Symbol SF(=).536 E F2(k).288
584: E/F4 6/Times-Roman@0 SF(0).344 1.6 M/F5 16/Symbol SF(S)-12.512 -8.934 M
585: /F6 11/Times-Italic@0 SF(g)5.772 -2.524 M F2(k).53 2.2 M/F7 11/Symbol SF
586: <b3>4.166 -2.2 M F1<a5>6.339 -9.454 M F2(k)-11.596 19.312 M F3(=).536 E
587: F2(k).288 E F4(0).344 1.6 M F5(S)-12.512 -8.934 M F0 -.165(ex)5.31
588: -2.524 O(p\().165 E F7(-)A F0(ln)1.87 E F6(k)2.266 E F0(\)).737 E F7(=)
589: 3.08 E F1<a5>6.339 -9.454 M F2(k)-11.596 19.312 M F3(=).536 E F2(k).288
590: E F4(0).344 1.6 M F5(S)-12.512 -8.934 M F0(1/)5.31 -2.524 M F6(k).396 E
591: F7(=)3.817 E/F8 14.3/Symbol SF<a5>3.08 1.716 M F0 235.701(.\()3.08
592: -1.716 O(11\))-235.701 E .198(Although there are sound ph)99.5 140.598 R
593: .197(ysical principles underlying the choices of Eqs. \(7\) and \(1\) \
594: \(Metropolis)-.055 F F6 .911(et al)72 162.598 R F0 3.661(,1)C .911
595: (953\), it w)-3.661 F .911
596: (as noted that this method of \214nding the global minimum in)-.11 F F6
597: (x)4.509 E F0 .912(-space w).583 F .912(as not limited to)-.11 F(ph)72
598: 184.598 Q 1.484(ysics e)-.055 F 1.484(xamples requiring)-.165 F F6 1.484
599: (bona \214de)4.234 F F0 1.484(\231temperatures\232 and \231ener)4.234 F
600: (gies.)-.198 E 6.983<9a52>-.77 G(ather)-6.983 E 4.233(,t)-.44 G 1.483
601: (his methodology can be)-4.233 F 1.881(readily e)72 206.598 R 1.881
602: (xtended to an)-.165 F 4.631(yp)-.165 G 1.881
603: (roblem for which a reasonable probability density)-4.631 F F6(h)4.973 E
604: F0(\().308 E F7(D)A F6(x).847 E F0 4.632(\)c).583 G 1.882
605: (an be formulated)-4.632 F(\(Kirkpatrick)72 228.598 Q F6(et al)2.75 E F0
606: 2.75(,1)C(983\).)-2.75 E/F9 11/Times-Bold@0 SF 2.75(2.1.2. Simulated)72
607: 272.598 R(quenching \(SQ\))2.75 E F0(Man)99.5 298.498 Q 5.726(yr)-.165 G
608: 2.976(esearchers ha)-5.726 F 3.306 -.165(ve f)-.22 H 2.976(ound it v)
609: .165 F 2.976(ery attracti)-.165 F 3.306 -.165(ve t)-.275 H 5.725(ot).165
610: G(ak)-5.725 E 5.725(ea)-.11 G(dv)-5.725 E 2.975
611: (antage of the ease of coding and)-.275 F .917
612: (implementing SA, utilizing its ability to handle quite comple)72
613: 320.498 R 3.667(xc)-.165 G .917(ost functions and constraints.)-3.667 F
614: (Ho)6.417 E(we)-.275 E -.165(ve)-.275 G -.44(r,).165 G .284
615: (the long time of e)72 342.498 R -.165(xe)-.165 G .284
616: (cution of standard Boltzmann-type SA has man).165 F 3.033(yt)-.165 G
617: .283(imes dri)-3.033 F -.165(ve)-.275 G 3.033(nt).165 G .283
618: (hese projects to utilize)-3.033 F 4.109(at)72 364.498 S 1.359
619: (emperature schedule too f)-4.109 F 1.359(ast to satisfy the suf)-.11 F
620: (\214cienc)-.275 E 4.109(yc)-.165 G 1.359
621: (onditions required to establish a true \(weak\))-4.109 F(er)72 386.498
622: Q .463(godic search.)-.198 F 3.213(Al)5.963 G(og)-3.213 E .462(arithmic\
623: temperature schedule is consistent with the Boltzmann algorithm, e.g.,\
624: the)-.055 F(temperature schedule is tak)72 408.498 Q(en to be)-.11 E F6
625: (T)107.901 442.861 Q F2(k).761 2.2 M F7(=)4.166 -2.2 M F6(T)2.981 E/F10
626: 8/Times-Roman@0 SF(0).473 2.2 M F0(ln)3.74 -9.9 M F6(k)2.266 E F10(0)
627: .473 2.2 M F0(ln)-18.038 13.2 M .44 LW 166.576 440.001 145.845 440.001
628: DL F6(k)158.812 450.561 Q F0 347.948(,\()7.28 -7.7 O(12\))-347.948 E
629: (where)72 475.332 Q F6(T)3.017 E F0 .367(is the \231temperature,)4.514 F
630: <9a>-.77 E F6(k)3.513 E F0 .367(is the \231time\232 inde)3.854 F 3.117
631: (xo)-.165 G 3.117(fa)-3.117 G .367(nnealing, and)-3.117 F F6(k)3.513 E
632: F10(0).473 2.2 M F0 .367(is some starting inde)3.667 -2.2 N 3.117
633: (x. This)-.165 F(can)3.117 E(be written for lar)72 497.332 Q(ge)-.198 E
634: F6(k)3.146 E F0(as)3.487 E F7(D)108 531.75 Q F6(T)-.099 E F7 3.08(=-)
635: 4.477 G F6(T)-3.179 E F10(0).473 2.2 M F0(ln)4.169 -9.9 M F6(k)2.266 E
636: F10(0).473 2.2 M F7(D).55 -2.2 M F6(k).396 E(k)-32.776 15.4 M F0(\(ln)
637: .737 E F6(k)2.266 E F0(\)).737 E 188.952 528.89 154.614 528.89 DL F10(2)
638: 184.402 536.37 Q F0(,)4.95 -4.62 M F6(k)3.476 E F0 4.73 -2.2(>> 1)3.817
639: H F6(T)107.901 578.41 Q F2(k).761 2.2 M F3(+).536 E F10(1)A F7(=)3.63
640: -2.2 M F6(T)2.981 E F2(k).761 2.2 M F7(-)3.506 -2.2 M F6(T)2.321 E F10
641: (0).473 2.2 M F0(ln)10.543 -9.9 M F6(k)2.266 E F10(0).473 2.2 M F6(k)
642: -26.588 13.2 M F0(\(ln).737 E F6(k)2.266 E F0(\)).737 E 210.87 575.55
643: 176.532 575.55 DL F10(2)206.32 583.03 Q F0 303.654(.\()4.95 -4.62 O
644: (13\))-303.654 E(Ho)72 612.707 Q(we)-.275 E -.165(ve)-.275 G .88 -.44
645: (r, s).165 H(ome researchers using the Boltzmann algorithm use e).44 E
646: (xponential schedules, e.g.,)-.165 E F6(T)107.901 641.207 Q F2(k).761
647: 2.2 M F3(+).536 E F10(1)A F7(=)3.63 -2.2 M F6(cT)3.3 E F2(k).761 2.2 M
648: F0 3.08(,0<)4.166 -2.2 O F6(c).22 E F0 3.08(<1)3.421 G F7(D)109.32
649: 674.425 Q F6(T)-.099 E(T)-11.483 15.4 M 123.466 679.265 109.32 679.265
650: DL F2(k)117.463 692.025 Q F7(=)6.851 -9.9 M F0(\()3.08 E F6(c).22 E F7
651: (-)2.761 E F0(1\))2.42 E F7(D)A F6(k).396 E F0(,)3.817 E F6(k)3.476 E F0
652: 4.73 -2.2(>> 1)3.817 H F6(T)107.901 723.263 Q F2(k).761 2.2 M F7(=)4.166
653: -2.2 M F6(T)2.981 E F10(0).473 2.2 M F0 -.165(ex)2.42 -2.2 O(p\().165 E
654: (\(\()-3.223 E F6(c).22 E F7(-)2.761 E F0(1\))2.42 E F6(k).396 E F0
655: 6.776 -3.223(\)\) ,).737 H(\(14\))315.45 E EP
656: %%Page: 7 7
657: %%BeginPageSetup
658: BP
659: %%EndPageSetup
660: /F0 11/Times-Roman@0 SF(ASA Lessons Learned)72 58 Q 165.888(-7- Lester)
661: 128.609 F(Ingber)2.75 E .691(with e)72 94 R(xpedienc)-.165 E 3.441(yt)
662: -.165 G .691(he only reason gi)-3.441 F -.165(ve)-.275 G 3.441(n. While)
663: .165 F .691(perhaps someday less stringent necessary conditions may)
664: 3.441 F .024(be de)72 116 R -.165(ve)-.275 G .024
665: (loped for the Boltzmann algorithm, this is not no).165 F 2.774(wt)-.275
666: G .024(he state of af)-2.774 F -.11(fa)-.275 G 2.774(irs. The).11 F .025
667: (question arises, what is)2.774 F 2.133(the v)72 138 R 2.133
668: (alue of this clear misuse of the claim to use SA to help solv)-.275 F
669: 4.883(et)-.165 G 2.133(hese problems/systems \(Ingber)-4.883 F(,)-.44 E
670: 3.805(1993b\)? Belo)72 160 R 2.485 -.715(w, a v)-.275 H 1.055
671: (ariant of SA, adapti).44 F 1.386 -.165(ve s)-.275 H 1.056
672: (imulated annealing \(ASA\) \(Ingber).165 F 3.806(,1)-.44 G 1.056
673: (989; Ingber)-3.806 F 3.806(,1)-.44 G 1.056(993a\), in)-3.806 F -.11(fa)
674: 72 182 S 1.398(ct does justify an e).11 F 1.397
675: (xponential annealing schedule, b)-.165 F 1.397
676: (ut only if a particular distrib)-.22 F 1.397(ution is used for the)-.22
677: F(generating function.)72 204 Q/F1 11/Times-Bold@0 SF 2.75(2.1.3. F)72
678: 248 R(ast annealing \(F)-.275 E(A\))-.99 E F0 .569
679: (Although there are man)99.5 273.9 R 3.319(yv)-.165 G .57
680: (ariants and impro)-3.594 F -.165(ve)-.165 G .57
681: (ments made on the \231standard\232 Boltzmann algorithm).165 F .322
682: (described abo)72 295.9 R -.165(ve)-.165 G 3.072(,m).165 G(an)-3.072 E
683: 3.072(yt)-.165 G -.165(ex)-3.072 G .321(tbooks \214nish just about at t\
684: his point without going into more detail about other).165 F .471
685: (algorithms that depart from this e)72 317.9 R .471
686: (xplicit algorithm \(v)-.165 F 2.75(an Laarho)-.275 F -.165(ve)-.165 G
687: 3.221(na).165 G .471(nd Aarts, 1987\).)-3.221 F(Speci\214cally)5.971 E
688: 3.221(,i)-.715 G 3.221(tw)-3.221 G(as)-3.331 E 2.372
689: (noted that the Cauch)72 339.9 R 5.122(yd)-.055 G(istrib)-5.122 E 2.372
690: (ution has some de\214nite adv)-.22 F 2.372(antages o)-.275 F -.165(ve)
691: -.165 G 5.121(rt).165 G 2.371(he Boltzmann form \(Szu and)-5.121 F
692: (Hartle)72 361.9 Q 1.43 -.715(y, 1)-.165 H 2.75(987\). The).715 F(Cauch)
693: 2.75 E 2.75(yd)-.055 G(istrib)-2.75 E(ution the)-.22 E 2.75(yd)-.165 G
694: (e\214ne is)-2.75 E/F2 11/Times-Italic@0 SF(g)108.462 395.933 Q F0(\()
695: .242 E/F3 11/Symbol SF(D)A F2(x).847 E F0(\)).583 E F3(=)3.08 E F2(T)
696: 37.955 -7.7 M F0(\()-39.671 15.4 M F3(D)A F2(x).847 E/F4 8/Times-Roman@0
697: SF(2).583 -3.08 M F3(+)2.97 3.08 M F2(T)2.321 E F4(2)1.397 -3.08 M F0
698: (\)).55 3.08 M F4(\()-3.08 I/F5 8/Times-Italic@0 SF(D).464 E/F6 8/Symbol
699: SF(+).224 E .44 LW 222.818 393.073 148.095 393.073 DL F4(1\)/2)209.38
700: 400.553 Q F0 291.706(,\()4.95 -4.62 O(15\))-291.706 E .266
701: (which has a \231f)72 430.23 R .266
702: (atter\232 tail than the Gaussian form of the Boltzmann distrib)-.11 F
703: .267(ution, permitting easier access to)-.22 F
704: (test local minima in the search for the desired global minimum.)72
705: 452.23 Q 1.279(It is instructi)99.5 478.13 R 1.608 -.165(ve t)-.275 H
706: 4.028(on).165 G 1.278
707: (ote the similar corresponding heuristic demonstration, that the Cauch)
708: -4.028 F(y)-.055 E F2(g)4.49 E F0(\().242 E F3(D)A F2(x).847 E F0(\))
709: .583 E(statistically \214nds a global minimum.)72 500.13 Q
710: (If Eq. \(8\) is replaced by)5.5 E F2(T)107.901 534.163 Q F0(\()1.397 E
711: F2(k).396 E F0(\)).737 E F3(=)3.08 E F2(T)4.301 -7.7 M F4(0).473 2.2 M
712: 153.316 531.303 142.276 531.303 DL F2(k)145.183 541.863 Q F0 361.208
713: (,\()7.649 -7.7 O(16\))-361.208 E(then here)72 566.634 Q/F7 10.4/Symbol
714: SF<a5>109.029 591.466 Q F5(k)-7.212 19.312 M/F8 6/Times-Roman@0 SF(0)
715: .344 1.6 M/F9 16/Symbol SF(S)-8.128 -8.934 M F2(g)2.332 -2.524 M F5(k)
716: .53 2.2 M F3<bb>4.166 -2.2 M F2(T)13.007 -7.7 M F4(0).473 2.2 M F3(D)
717: -19.196 13.2 M F2(x).847 E F5(D)1.047 -3.08 M F6(+).224 E 172.443 598.06
718: 143.991 598.06 DL F4(1)167.893 605.54 Q F7<a5>4.769 -14.074 M F5(k)
719: -7.212 19.312 M F8(0).344 1.6 M F9(S)-8.128 -8.934 M F0(1)3.448 -10.224
720: M 194.312 598.06 188.295 598.06 DL F2(k)188.691 608.62 Q F3(=)5.137 -7.7
721: M/F10 14.3/Symbol SF<a5>3.08 1.716 M F0 297.817(.\()3.08 -1.716 O(17\))
722: -297.817 E .839(Note that the \231normalization\232 of)72 638.212 R F2
723: (g)4.051 E F0 .839(has introduced the annealing-time inde)3.831 F(x)
724: -.165 E F2(k)3.985 E F0 3.59(,g).737 G -.275(iv)-3.59 G .84
725: (ing some insights into).275 F(ho)72 660.212 Q 4.685(wt)-.275 G 4.685
726: (oc)-4.685 G 1.935(onstruct other annealing distrib)-4.685 F 4.685
727: (utions. The)-.22 F 1.935(method of F)4.685 F 4.685(Ai)-.814 G 4.685(st)
728: -4.685 G 1.935(hus seen to ha)-4.685 F 2.265 -.165(ve a)-.22 H 4.684(na)
729: .165 G(nnealing)-4.684 E 2.5(schedule e)72 682.212 R 2.5(xponentially f)
730: -.165 F 2.5(aster than the method of B)-.11 F 5.25(A. This)-.385 F 2.5
731: (method has been tested in a v)5.25 F 2.501(ariety of)-.275 F
732: (problems \(Szu and Hartle)72 704.212 Q 1.43 -.715(y, 1)-.165 H(987\).)
733: .715 E EP
734: %%Page: 8 8
735: %%BeginPageSetup
736: BP
737: %%EndPageSetup
738: /F0 11/Times-Roman@0 SF(ASA Lessons Learned)72 58 Q 165.888(-8- Lester)
739: 128.609 F(Ingber)2.75 E/F1 11/Times-Bold@0 SF 2.75(2.2. Adapti)72 94 R
740: .22 -.11(ve S)-.11 H(imulated Annealing \(ASA\)).11 E F0 .533(In a v)
741: 99.5 119.9 R .533(ariety of ph)-.275 F .533(ysical problems we ha)-.055
742: F .863 -.165(ve a)-.22 H/F2 11/Times-Italic@0 SF(D)4.086 E F0 .533
743: (-dimensional parameter).308 F 3.283(-space. Dif)-.22 F .533
744: (ferent parameters)-.275 F(ha)72 141.9 Q 2.765 -.165(ve d)-.22 H(if).165
745: E 2.436(ferent \214nite ranges, \214x)-.275 F 2.436(ed by ph)-.165 F
746: 2.436(ysical considerations, and dif)-.055 F 2.436
747: (ferent annealing-time-dependent)-.275 F(sensiti)72 163.9 Q 2.985
748: (vities, measured by the curv)-.275 F 2.984
749: (ature of the cost-function at local minima.)-.275 F 3.754 -.385(BA a)
750: 8.484 H 2.984(nd F).385 F 5.734(Ah)-.814 G -2.475 -.22(av e)-5.734 H F2
751: (g)6.416 E F0(distrib)72 185.9 Q 1.07
752: (utions which sample in\214nite ranges, and there is no pro)-.22 F 1.07
753: (vision for considering dif)-.165 F 1.07(ferences in each)-.275 F
754: (parameter)72 207.9 Q 1.143(-dimension, e.g., dif)-.22 F 1.143
755: (ferent sensiti)-.275 F 1.142(vities might require dif)-.275 F 1.142
756: (ferent annealing schedules.)-.275 F 1.142(These are)6.642 F 1.672
757: (among se)72 229.9 R -.165(ve)-.275 G 1.672(ral considerations that g)
758: .165 F -2.475 -.22(av e)-.055 H 1.672(rise to Adapti)4.642 F 2.002 -.165
759: (ve S)-.275 H 1.673(imulated Annealing \(ASA\).).165 F 1.673
760: (Full details are)7.173 F -.22(av)72 251.9 S
761: (ailable by obtaining the publicly a)-.055 E -.275(va)-.22 G
762: (ilable source code \(Ingber).275 E 2.75(,1)-.44 G(993a\).)-2.75 E
763: (ASA considers a parameter)99.5 278.302 Q/F3 11/Symbol-Slanted SF(a)
764: 2.376 E/F4 8/Times-Italic@0 SF(i)1.614 -4.62 M(k)-2.461 7.312 M F0
765: (in dimension)3.836 -2.692 M F2(i)2.761 E F0
766: (generated at annealing-time)3.146 E F2(k)3.146 E F0(with the range)
767: 3.487 E F3(a)107.626 307.304 Q F4(i)1.614 -4.62 M(k)-2.461 7.312 M/F5 11
768: /Symbol SF<ce>1.966 -2.692 M F0([)-.88 E F2(A)1.111 E F4(i).041 2.2 M F0
769: (,).838 -2.2 M F2(B)2.508 E F4(i).305 2.2 M F0 345.661 3.08(], \().838
770: -2.2 P(18\))-3.08 E(calculated with the random v)72 336.306 Q(ariable)
771: -.275 E F2(y)3.564 E F4(i).36 -4.62 M F0(,).838 4.62 M F3(a)107.626
772: 365.308 Q F4(i)1.614 -4.62 M(k)-2.461 7.312 M/F6 8/Symbol SF(+).536 E/F7
773: 8/Times-Roman@0 SF(1)A F5(=)3.63 -2.692 M F3(a)2.706 E F4(i)1.614 -4.62
774: M(k)-2.461 7.312 M F5(+)3.506 -2.692 M F2(y)3.234 E F4(i).36 -4.62 M F0
775: (\().838 4.62 M F2(B).638 E F4(i).305 2.2 M F5(-)3.258 -2.2 M F2(A)3.531
776: E F4(i).041 2.2 M F0 3.08(\),).838 -2.2 O F2(y)108.814 400.81 Q F4(i).36
777: -4.62 M F5<ce>1.718 4.62 M F0([)-.88 E F5(-)A F0 -.88(1, 1])B 361.896
778: (.\()3.08 G(19\))-361.896 E(De\214ne the generating function)99.5 433.21
779: Q F2(g)108.462 469.128 Q F4(T).17 2.2 M F0(\()1.566 -2.2 M F2(y).814 E
780: F0(\)).352 E F5(=)3.08 E F4(D)5.246 -10.444 M(i)-8.456 20.246 M F6(=)
781: .288 E F7(1)A/F8 16/Symbol SF(P)-11.6 -7.278 M F0(1)49.958 -10.224 M
782: (2\(|)-52.268 15.4 M F2(y).814 E F4(i).36 -3.08 M F0(|).838 3.08 M F5(+)
783: 2.42 E F2(T)2.321 E F4(i).481 2.2 M F0 1.87(\)l).838 -2.2 O(n\(1)-1.87 E
784: F5(+)2.42 E F0(1/)2.42 E F2(T)-.099 E F4(i).481 2.2 M .44 LW 259.025
785: 466.268 159.989 466.268 DL F0(\))255.362 476.828 Q F5<ba>4.4 -7.7 M F4
786: (D)5.246 -10.444 M(i)-8.456 20.246 M F6(=).288 E F7(1)A F8(P)-11.6
787: -7.278 M F2(g)2.332 -2.524 M F4(i).25 -4.62 M(T)-2.304 7.15 M F0(\()
788: 1.566 -2.53 M F2(y).814 E F4(i).36 -4.62 M F0 192.85 3.08(\), \().838
789: 4.62 P(20\))-3.08 E .203(where the subscript)72 504.768 R F2(i)2.963 E
790: F0(on)3.348 E F2(T)2.853 E F4(i).481 2.2 M F0 .202
791: (speci\214es the parameter inde)3.79 -2.2 N .202(x, and the)-.165 F F2
792: (k)3.348 E F0 .202(-dependence in).737 F F2(T)2.853 E F4(i).481 2.2 M F0
793: (\().838 -2.2 M F2(k).396 E F0 2.952(\)f).737 G .202(or the annealing)
794: -2.952 F(schedule has been dropped for bre)72 526.768 Q(vity)-.275 E 5.5
795: (.I)-.715 G(ts cumulati)-5.5 E .33 -.165(ve p)-.275 H
796: (robability distrib).165 E(ution is)-.22 E F2(G)107.978 569.471 Q F4(T)
797: .401 2.2 M F0(\()1.566 -2.2 M F2(y).814 E F0(\)).352 E F5(=)3.08 E F4(y)
798: 4.221 -15.437 M/F9 6/Times-Roman@0 SF(1).256 -2.96 M F6(-)-9.159 32.533
799: M F7(1)A/F10 19/Symbol SF<f2>-6.21 -9.311 M F0 1.87(...)2.874 -7.196 O
800: F4(y)1.427 -13.066 M/F11 6/Times-Italic@0 SF(D).604 -2.96 M F6(-)-9.915
801: 32.533 M F7(1)A F10<f2>-6.21 -9.311 M F2(dy)4.515 -4.825 M F5<a2>.352 E
802: F7(1)-4.62 I F0 1.87(...)2.42 2.249 O F2(dy).385 2.371 M F5<a2>.352 E F4
803: (D).464 -4.62 M F2(g)3.106 4.62 M F4(T).17 2.2 M F0(\()1.566 -2.2 M F2
804: (y).814 E F5<a2>.352 E F0(\))A F5<ba>3.08 E F4(D)5.246 -10.444 M(i)
805: -8.456 20.246 M F6(=).288 E F7(1)A F8(P)-11.6 -7.278 M F2(G)1.848 -2.524
806: M F4(i).558 -4.62 M(T)-2.381 7.15 M F0(\()1.566 -2.53 M F2(y).814 E F4
807: (i).36 -4.62 M F0 3.08(\),).838 4.62 O F2(G)107.978 624.059 Q F4(i).558
808: -4.62 M(T)-2.381 7.15 M F0(\()1.566 -2.53 M F2(y).814 E F4(i).36 -4.62 M
809: F0(\)).838 4.62 M F5(=)3.08 E F0(1)4.4 -7.7 M 157.8 621.199 152.3
810: 621.199 DL(2)152.3 631.759 Q F5(+)3.74 -7.7 M F0 -.88(sgn \()5.61 -7.7 N
811: F2(y).814 E F4(i).36 -4.62 M F0(\)).838 4.62 M 206.784 621.199 171.319
812: 621.199 DL(2)186.301 631.759 Q(ln\(1)19.493 -15.4 M F5(+)2.42 E F0(|)
813: 2.42 E F2(y).814 E F4(i).36 -4.62 M F0(|/).838 4.62 M F2(T)-.099 E F4(i)
814: .481 2.2 M F0(\)).838 -2.2 M(ln\(1)-54.391 15.4 M F5(+)2.42 E F0(1/)2.42
815: E F2(T)-.099 E F4(i).481 2.2 M 269.695 621.199 211.294 621.199 DL F0(\))
816: 262.022 631.759 Q 244.829(.\()8.41 -7.7 O(21\))-244.829 E F2(y)72.814
817: 659.199 Q F4(i).36 -4.62 M F0(is generated from a)3.588 4.62 M F2(u)
818: 2.838 E F4(i).283 -4.62 M F0(from the uniform distrib)3.588 4.62 M
819: (ution)-.22 E F2(u)108.088 688.201 Q F4(i).283 -4.62 M F5<ce>1.718 4.62
820: M F2(U)-1.452 E F0 -.88([0, 1])1.023 F(,)3.08 E F2(y)108.814 727.635 Q
821: F4(i).36 -4.62 M F5(=)3.918 4.62 M F0 -.88(sgn \()1.87 F F2(u).088 E F4
822: (i).283 -4.62 M F5(-)3.258 4.62 M/F12 9/Times-Roman@0 SF(1)3.74 -7.7 M
823: 174.553 724.775 170.053 724.775 DL(2)170.053 735.335 Q F0(\))1.32 -7.7 M
824: F2(T)-.099 E F4(i).481 2.2 M F0([\(1).838 -2.2 M F5(+)2.42 E F0(1/)2.42
825: E F2(T)-.099 E F4(i).481 2.2 M F0(\)).838 -2.2 M F7(|2)-4.62 I F4(u).064
826: E F11(i).206 -2.96 M F6(-).616 2.96 M F7(1|)A F5(-)2.97 4.62 M F0 .33
827: (1] .)2.42 F(\(22\))238.524 E EP
828: %%Page: 9 9
829: %%BeginPageSetup
830: BP
831: %%EndPageSetup
832: /F0 11/Times-Roman@0 SF(ASA Lessons Learned)72 58 Q 165.888(-9- Lester)
833: 128.609 F(Ingber)2.75 E(It is straightforw)99.5 94 Q
834: (ard to calculate that for an annealing schedule for)-.11 E/F1 11
835: /Times-Italic@0 SF(T)2.651 E/F2 8/Times-Italic@0 SF(i).481 2.2 M F1(T)
836: 107.901 123.178 Q F2(i).481 2.2 M F0(\().838 -2.2 M F1(k).396 E F0(\))
837: .737 E/F3 11/Symbol SF(=)3.08 E F1(T)2.981 E/F4 8/Times-Roman@0 SF(0)
838: .473 2.2 M F2(i).008 E F0 -.165(ex)2.708 -2.2 O(p\().165 E F3(-)A F1(c)
839: .22 E F2(i).349 2.2 M F1(k)1.234 -2.2 M F4(1/).737 -4.62 M F2(D).464 E
840: F0 291.548 3.08(\), \().774 4.62 P(23\))-3.08 E 2.75(ag)72 151.678 S
841: (lobal minima statistically can be obtained.)-2.75 E(I.e.,)5.5 E/F5 10.4
842: /Symbol SF<a5>109.029 178.142 Q F2(k)-7.212 19.312 M/F6 6/Times-Roman@0
843: SF(0).344 1.6 M/F7 16/Symbol SF(S)-8.128 -8.934 M F1(g)2.332 -2.524 M F2
844: (k).53 2.2 M F3<bb>4.166 -2.2 M F5<a5>2.899 -9.454 M F2(k)-7.212 19.312
845: M F6(0).344 1.6 M F7(S)-8.128 -8.934 M F0([)1.87 -2.524 M F2(D)5.246
846: -10.444 M(i)-8.456 20.246 M/F8 8/Symbol SF(=).288 E F4(1)A F7(P)-11.6
847: -7.278 M F0(1)14.207 -10.224 M(2|)-16.517 15.4 M F1(y).814 E F2(i).36
848: -3.08 M F0(|).838 3.08 M F1(c).22 E .44 LW 201.349 184.736 173.814
849: 184.736 DL F2(i)198.287 197.496 Q F0(])2.158 -9.9 M(1)3.448 -7.7 M
850: 215.539 184.736 209.522 184.736 DL F1(k)209.918 195.296 Q F3(=)5.137
851: -7.7 M/F9 14.3/Symbol SF<a5>3.08 1.716 M F0 276.59(.\()3.08 -1.716 O
852: (24\))-276.59 E(It seems sensible to choose control o)72 224.888 Q -.165
853: (ve)-.165 G(r).165 E F1(c)2.97 E F2(i).349 2.2 M F0 2.75(,s).838 -2.2 O
854: (uch that)-2.75 E F1(T)107.901 253.388 Q F2(fi)2.049 2.2 M F3(=)3.918
855: -2.2 M F1(T)2.981 E F4(0).473 2.2 M F2(i).008 E F0 -.165(ex)2.708 -2.2 O
856: (p\().165 E F3(-)A F1(m).418 E F2(i).36 2.2 M F0 3.08(\)w).838 -2.2 O
857: (hen)-3.08 E F1(k)3.476 E F2(f)2.049 2.2 M F3(=)5.198 -2.2 M F0 -.165
858: (ex)3.08 G(p).165 E F1(n)2.266 E F2(i).272 2.2 M F0(,)3.918 -2.2 M F1(c)
859: 108.22 288.394 Q F2(i).349 2.2 M F3(=)3.918 -2.2 M F1(m)3.498 E F2(i).36
860: 2.2 M F0 -.165(ex)2.708 -2.2 O(p\().165 E F3(-)A F1(n).396 E F2(i).272
861: 2.2 M F0(/).838 -2.2 M F1(D).638 E F0 314.308 3.08(\), \().308 H(25\))
862: -3.08 E(where)72 316.894 Q F1(m)3.168 E F2(i).36 2.2 M F0(and)3.588 -2.2
863: M F1(n)3.146 E F2(i).272 2.2 M F0(can be considered \231free\232 parame\
864: ters to help tune ASA for speci\214c problems.)3.588 -2.2 M .933
865: (It has pro)99.5 342.794 R -.165(ve)-.165 G 3.683(nf).165 G .933(ruitfu\
866: l to use the same type of annealing schedule for the acceptance functio\
867: n)-3.683 F F1(h)4.025 E F0(as)3.992 E 2.012
868: (used for the generating function)72 364.794 R F1(g)5.224 E F0 4.762(,i)
869: .242 G 2.012(.e., Equations \(23\) and \(25\), b)-4.762 F 2.012
870: (ut with the number of acceptance)-.22 F .215(points, instead of the nu\
871: mber of generated points, used to determine the)72 386.794 R F1(k)3.362
872: E F0 .216(for the acceptance temperature.)3.703 F 1.05
873: (In one implementation of this algorithm, ne)99.5 413.196 R 3.8(wp)-.275
874: G(arameters)-3.8 E/F10 11/Symbol-Slanted SF(a)3.426 E F2(i)1.614 -4.62 M
875: (k)-2.461 7.312 M F8(+).536 E F4(1)A F0 1.05
876: (are generated from old parameters)4.35 -2.692 N F10(a)71.626 435.698 Q
877: F2(i)1.614 -4.62 M(k)-2.461 7.312 M F0 2.322(by generating the)6.158
878: -2.692 N F1(y)5.886 E F2(i).36 -4.62 M F0 3.532 -.605('s u).838 4.62 P
879: 2.323(ntil a set of).605 F F1(D)5.711 E F0 2.323
880: (are obtained satisfying the range constraints.)5.381 F 2.323
881: (In another)7.823 F(alternati)72 458.2 Q .669 -.165(ve s)-.275 H .338
882: (upported in ASA, useful for some constraint problems, the).165 F F1(y)
883: 3.902 E F2(i).36 -4.62 M F0 1.548 -.605('s a).838 4.62 P .338
884: (re generated sequentially for).605 F(each test of the cost function.)72
885: 480.2 Q/F11 11/Times-Bold@0 SF 2.75(2.2.1. Reannealing)72 524.2 R F0
886: (Whene)99.5 550.1 Q -.165(ve)-.275 G 5.701(rd).165 G 2.951
887: (oing a multi-dimensional search in the course of a real-w)-5.701 F
888: 2.952(orld nonlinear ph)-.11 F(ysical)-.055 E 1.294(problem, ine)72
889: 572.602 R 1.294(vitably one must deal with dif)-.275 F 1.294
890: (ferent changing sensiti)-.275 F 1.293(vities of the)-.275 F F10(a)3.669
891: E F2(i)1.614 -4.62 M F0 1.293(in the search.)4.881 4.62 N 1.293(At an)
892: 6.793 F(y)-.165 E(gi)72 594.602 Q -.165(ve)-.275 G 4.335(na).165 G 1.585
893: (nnealing-time, it seems sensible to attempt to \231stretch out\232 the\
894: range o)-4.335 F -.165(ve)-.165 G 4.335(rw).165 G 1.585
895: (hich the relati)-4.335 F -.165(ve)-.275 G(ly).165 E(insensiti)72
896: 616.602 Q .33 -.165(ve p)-.275 H(arameters are being searched, relati)
897: .165 E .33 -.165(ve t)-.275 H 2.75(ot).165 G
898: (he ranges of the more sensiti)-2.75 E .33 -.165(ve p)-.275 H
899: (arameters.).165 E .734(It has pro)99.5 642.502 R -.165(ve)-.165 G 3.483
900: (nf).165 G .733(ruitful to accomplish this by periodically rescaling th\
901: e annealing-time)-3.483 F F1(k)3.879 E F0 3.483(,e).737 G(ssentially)
902: -3.483 E 1.105(reannealing, e)72 664.502 R -.165(ve)-.275 G 1.105
903: (ry hundred or so acceptance-e).165 F -.165(ve)-.275 G 1.105
904: (nts \(or at some user).165 F 1.105(-de\214ned modulus of the number of)
905: -.22 F .161(accepted or generated states\), in terms of the sensiti)72
906: 686.502 R(vities)-.275 E F1(s)3.285 E F2(i).305 2.2 M F0 .161
907: (calculated at the most current minimum v)3.749 -2.2 N(alue)-.275 E
908: 168.112 710.042 164.702 710.042 DL(of the cost function,)72 708.502 Q F1
909: (L)3.388 E F0(,).582 E F1(s)108.374 737.504 Q F2(i).305 2.2 M 139.026
910: 739.044 135.616 739.044 DL F3 3.08<3db6>119.1 737.504 S F1(L)-2.442 E F0
911: (/).582 E F3<b6>A F10(a)-.374 E F2(i)1.614 -4.62 M F0 355.879(.\()3.918
912: 4.62 O(26\))-355.879 E EP
913: %%Page: 10 10
914: %%BeginPageSetup
915: BP
916: %%EndPageSetup
917: /F0 11/Times-Roman@0 SF(ASA Lessons Learned)72 58 Q 163.138(-10- Lester)
918: 125.859 F(Ingber)2.75 E .121(In terms of the lar)72 94 R(gest)-.198 E/F1
919: 11/Times-Italic@0 SF(s)3.245 E/F2 8/Times-Italic@0 SF(i).305 2.2 M F0(=)
920: 3.709 -2.2 M F1(s)3.245 E/F3 8/Times-Roman@0 SF(max).297 2.2 M F0 2.871
921: (,ad).55 -2.2 O(ef)-2.871 E .121(ault rescaling is performed for each)
922: -.11 F F1(k)3.267 E F2(i).481 2.2 M F0 .121
923: (of each parameter dimension,)3.709 -2.2 N(whereby a ne)72 116 Q 2.75
924: (wi)-.275 G(nde)-2.75 E(x)-.165 E F1(k)3.146 E/F4 11/Symbol SF<a2>.737 E
925: F2(i).008 2.2 M F0(is calculated from each)3.588 -2.2 M F1(k)3.146 E F2
926: (i).481 2.2 M F0(,).838 -2.2 M F1(k)108.396 144.5 Q F2(i).481 2.2 M F4
927: <ae>3.918 -2.2 M F1(k)3.476 E F4<a2>.737 E F2(i).008 2.2 M F0(,)3.918
928: -2.2 M F1(T)107.901 179.5 Q F4<a2>1.397 E F2(ik).008 2.2 M/F5 8/Symbol
929: SF<a2>.536 E F4(=)3.63 -2.2 M F1(T)2.981 E F2(ik).481 2.2 M F0(\()1.086
930: -2.2 M F1(s).374 E F3(max).297 2.2 M F0(/).55 -2.2 M F1(s).374 E F2(i)
931: .305 2.2 M F0 3.08(\),).838 -2.2 O F1(k)108.396 214.994 Q F4<a2>.737 E
932: F2(i).008 2.2 M F4(=)3.918 -2.2 M F0 -3.223(\(\()3.08 G(ln\()3.223 E F1
933: (T)-.099 E F2(i).481 2.2 M F3(0).288 E F0(/).55 -2.2 M F1(T)-.099 E F2
934: (ik).481 2.2 M F5<a2>.536 E F0(\)/).55 -2.2 M F1(c).22 E F2(i).349 2.2 M
935: F0 -3.223(\)\)).838 -2.2 O F2(D)3.687 -4.62 M F0 309.21(.\()3.854 4.62 O
936: (27\))-309.21 E F1(T)71.901 243.494 Q F2(i).481 2.2 M F3(0).288 E F0
937: (is set to unity to be)3.3 -2.2 M
938: (gin the search, which is ample to span each parameter dimension.)-.165
939: E 1.477(The acceptance temperature is similarly rescaled.)99.5 269.394 R
940: 1.477(Since the initial acceptance temperature is set)6.977 F .423
941: (equal to an initial trial v)72 291.394 R .44 LW 218.46 292.934 215.05
942: 292.934 DL .423(alue of)179.252 291.394 R F1(L)3.811 E F0 3.173(,t).582
943: G .423(his is typically v)-3.173 F .423(ery lar)-.165 F .423(ge relati)
944: -.198 F .753 -.165(ve t)-.275 H 3.173(ot).165 G .423
945: (he current best minimum, which)-3.173 F 1.517
946: (may tend to distort the scale of the re)72 313.394 R 1.517
947: (gion currently being sampled.)-.165 F 1.517
948: (Therefore, when this rescaling is)7.017 F .649(performed, the initial \
949: acceptance temperature is reset to the maximum of the most current mini\
950: mum and)72 335.394 R 208.576 358.934 205.166 358.934 DL .637
951: (the best current minimum of)72 357.394 R F1(L)4.025 E F0 3.387(,a).582
952: G .637(nd the annealing-time inde)-3.387 F 3.387(xa)-.165 G .637
953: (ssociated with this temperature is reset to)-3.387 F(gi)72 379.394 Q
954: .943 -.165(ve a n)-.275 H 1.163 -.275(ew t).165 H .614(emperature equal\
955: to the minimum of the current cost-function and the absolute v).275 F
956: .614(alues of the)-.275 F(current best and last minima.)72 401.394 Q
957: .041(Also generated are the \231standard de)99.5 427.972 R .041
958: (viations\232 of the theoretical forms, calculated as [)-.275 F F4<b6>A
959: 485.606 429.512 482.196 429.512 DL F3(2)475.683 423.352 Q F1(L)1.188
960: 4.62 M F0(/\().582 E F4<b6>A/F6 11/Symbol-Slanted SF(a)-.374 E F2(i)
961: 1.614 -4.62 M F0(\)).838 4.62 M F3(2)-4.62 I F0(]).55 4.62 M F5(-)-4.62
962: I F3(1/2)A F0(,).55 4.62 M 1.038(for each parameter)72 449.972 R F6(a)
963: 3.414 E F2(i)1.097 2.2 M F0 6.538(.T).838 -2.2 O 1.038(his gi)-6.538 F
964: -.165(ve)-.275 G 3.788(sa).165 G 3.788(ne)-3.788 G 1.038(stimate of the\
965: \231noise\232 that accompanies \214ts to stochastic data or)-3.788 F
966: 3.464(functions. At)72 471.972 R .714(the end of the run, the of)3.464 F
967: .714(f-diagonal elements of the \231co)-.275 F -.275(va)-.165 G .714
968: (riance matrix\232 are calculated for).275 F 3.615(all parameters.)72
969: 493.972 R 3.615(This in)9.115 F -.165(ve)-.44 G 3.615(rse curv).165 F
970: 3.615(ature of the theoretical cost function can pro)-.275 F 3.616
971: (vide a quantitati)-.165 F -.165(ve)-.275 G(assessment of the relati)72
972: 515.972 Q .33 -.165(ve s)-.275 H(ensiti).165 E(vity of parameters to st\
973: atistical errors in \214ts to stochastic systems.)-.275 E/F7 11
974: /Times-Bold@0 SF 2.75(2.2.2. Quenching)72 559.972 R F0 .624
975: (Another adapti)99.5 585.872 R .954 -.165(ve f)-.275 H .624
976: (eature of ASA is its ability to perform quenching in a methodical f)
977: .165 F 3.373(ashion. This)-.11 F
978: (is applied by noting that the temperature schedule abo)72 607.872 Q .33
979: -.165(ve c)-.165 H(an be rede\214ned as).165 E F1(T)107.901 638.056 Q F2
980: (i).481 2.2 M F0(\().838 -2.2 M F1(k).396 E F2(i).481 2.2 M F0(\)).838
981: -2.2 M F4(=)3.08 E F1(T)2.981 E F3(0).473 2.2 M F2(i).008 E F0 -.165(ex)
982: 2.708 -2.2 O(p\().165 E F4(-)A F1(c).22 E F2(i).349 2.2 M F1(k)1.234
983: -2.2 M F2(Q).665 -5.626 M/F8 6/Times-Italic@0 SF(i).222 1.6 M F3(/).616
984: -1.6 M F2(D).464 E(i)-16.93 8.658 M F0 3.08(\),)15.48 -3.032 O F1(c)
985: 108.22 673.056 Q F2(i).349 2.2 M F4(=)3.918 -2.2 M F1(m)3.498 E F2(i).36
986: 2.2 M F0 -.165(ex)2.708 -2.2 O(p\().165 E F4(-)A F1(n).396 E F2(i).272
987: 2.2 M F1(Q).739 -2.2 M F2(i).305 2.2 M F0(/).838 -2.2 M F1(D).638 E F0
988: 303.098 3.08(\), \().308 H(28\))-3.08 E(in terms of the \231quenching f)
989: 72 701.556 Q(actor\232)-.11 E F1(Q)2.651 E F2(i).305 2.2 M F0 5.5(.T)
990: .838 -2.2 O(he abo)-5.5 E .33 -.165(ve p)-.165 H(roof f).165 E
991: (ails at Eq. \(24\) if)-.11 E F1(Q)2.651 E F2(i).305 2.2 M F0 3.08(>1)
992: 3.918 -2.2 O(as)-.33 E EP
993: %%Page: 11 11
994: %%BeginPageSetup
995: BP
996: %%EndPageSetup
997: /F0 11/Times-Roman@0 SF(ASA Lessons Learned)72 58 Q 163.138(-11- Lester)
998: 125.859 F(Ingber)2.75 E/F1 8/Times-Italic@0 SF(k)110.836 111.276 Q/F2 16
999: /Symbol SF(S)-6.388 -7.334 M F1(D)5.246 -12.968 M F2(P)-9.152 12.968 M
1000: F0(1/)1.87 -2.524 M/F3 11/Times-Italic@0 SF(k).396 E F1(Q).665 -4.62 M
1001: /F4 6/Times-Italic@0 SF(i).222 1.6 M/F5 8/Times-Roman@0 SF(/).616 -1.6 M
1002: F1(D).464 E/F6 11/Symbol SF(=)3.854 4.62 M F1(k)5.916 9.858 M F2(S)
1003: -6.388 -7.334 M F0(1/)1.87 -2.524 M F3(k).396 E F1(Q).665 -4.62 M F4(i)
1004: .222 1.6 M F0(<)4.246 3.02 M/F7 14.3/Symbol SF<a5>3.08 1.716 M F0
1005: 280.885(.\()3.08 -1.716 O(29\))-280.885 E .917
1006: (This simple calculation sho)99.5 141.014 R .917(ws ho)-.275 F 3.667(wt)
1007: -.275 G .917(he \231curse of dimensionality\232 arises, and also gi)
1008: -3.667 F -.165(ve)-.275 G 3.668(sap).165 G(ossible)-3.668 E -.11(wa)72
1009: 163.014 S 3.491(yo).11 G 3.49(fl)-3.491 G -.275(iv)-3.49 G .74
1010: (ing with this disease.).275 F .74(In ASA, the in\215uence of lar)6.24 F
1011: .74(ge dimensions becomes clearly focussed on)-.198 F 1.771(the e)72
1012: 185.014 R 1.772(xponential of the po)-.165 F 1.772(wer of)-.275 F F3(k)
1013: 4.918 E F0 1.772(being 1/)5.259 F F3(D).638 E F0 4.522(,a).308 G 4.522
1014: (st)-4.522 G 1.772(he annealing required to properly sample the space)
1015: -4.522 F .339(becomes prohibiti)72 207.014 R -.165(ve)-.275 G .339
1016: (ly slo).165 F 4.519 -.715(w. S)-.275 H .339
1017: (o, if we cannot commit resources to properly sample the space er).715 F
1018: (godically)-.198 E(,)-.715 E .901(then for some systems perhaps the ne)
1019: 72 229.014 R .901(xt best procedure w)-.165 F .901
1020: (ould be to turn on quenching, whereby)-.11 F F3(Q)3.552 E F1(i).305 2.2
1021: M F0(can)4.489 -2.2 M
1022: (become on the order of the size of number of dimensions.)72 251.014 Q
1023: .19(The scale of the po)99.5 276.914 R .19(wer of 1/)-.275 F F3(D).638 E
1024: F0 .189
1025: (temperature schedule used for the acceptance function can be altered)
1026: 3.247 F 1.407(in a similar f)72 298.914 R 4.157(ashion. Ho)-.11 F(we)
1027: -.275 E -.165(ve)-.275 G 2.287 -.44(r, t).165 H 1.407(his does not af)
1028: .44 F 1.407(fect the annealing proof of ASA, and so this may used)-.275
1029: F(without damaging the \(weak\) er)72 320.914 Q(godicity property)-.198
1030: E(.)-.715 E/F8 11/Times-Bold@0 SF 2.75(2.2.3. ASA)72 364.914 R
1031: (applications)2.75 E F0 .804(The abo)99.5 390.814 R 1.134 -.165(ve d)
1032: -.165 H .804(e\214nes this method of adapti).165 F 1.133 -.165(ve s)
1033: -.275 H .803(imulated annealing \(ASA\), pre).165 F .803
1034: (viously called v)-.275 F .803(ery f)-.165 F(ast)-.11 E .662
1035: (simulated reannealing \(VFSR\) \(Ingber)72 412.814 R 3.412(,1)-.44 G
1036: .663(989\) only named such to contrast it the pre)-3.412 F .663
1037: (vious method of f)-.275 F(ast)-.11 E 2.593(annealing \(F)72 434.814 R
1038: 2.592(A\) \(Szu and Hartle)-.814 F 4.022 -.715(y, 1)-.165 H 5.342
1039: (987\). The).715 F 2.592(annealing schedules for the temperatures)5.342
1040: F F3(T)5.243 E F1(i).481 2.2 M F0(decrease)6.18 -2.2 M -.165(ex)72
1041: 457.492 S 3.366(ponentially in annealing-time).165 F F3(k)6.513 E F0
1042: 6.117(,i).737 G(.e.,)-6.117 E F3(T)6.018 E F1(i).481 2.2 M F6(=)3.918
1043: -2.2 M F3(T)2.981 E F1(i).481 2.2 M F5(0).288 E F0 -.165(ex)2.42 -2.2 O
1044: (p\().165 E F6(-)A F3(c).22 E F1(i).349 2.2 M F3(k)1.234 -2.2 M F5(1/)
1045: .737 -4.62 M F1(D).464 E F0 6.117(\). Of).774 4.62 N 3.367
1046: (course, the f)6.117 F 3.367(atter the tail of the)-.11 F 2.581(generat\
1047: ing function, the smaller the ratio of acceptance to generated points i\
1048: n the \214t.)72 479.492 R(Ho)8.081 E(we)-.275 E -.165(ve)-.275 G 3.46
1049: -.44(r, i).165 H(n).44 E .054(practice, it is found that for a gi)72
1050: 501.492 R -.165(ve)-.275 G 2.805(ng).165 G .055(enerating function, thi\
1051: s ratio is approximately constant as the \214t \214nds)-2.805 F 3.596
1052: (ag)72 523.492 S .846(lobal minimum.)-3.596 F .846(Therefore, for a lar)
1053: 6.346 F .846(ge parameter space, the ef)-.198 F(\214cienc)-.275 E 3.596
1054: (yo)-.165 G 3.596(ft)-3.596 G .846(he \214t is determined by the)-3.596
1055: F(annealing schedule of the generating function.)72 545.492 Q 3.19(Am)
1056: 99.5 571.392 S .44(ajor dif)-3.19 F .44(ference between ASA and B)-.275
1057: F 3.19(Aa)-.385 G .44(lgorithms is that the er)-3.19 F .441
1058: (godic sampling tak)-.198 F .441(es place in an)-.11 F F3(n)72.396
1059: 593.392 Q F6(+)2.684 E F0 4.173(1d)2.42 G 1.422
1060: (imensional space, i.e., in terms of)-4.173 F F3(n)4.568 E F0 1.422
1061: (parameters and the cost function.)4.436 F 1.422(In ASA the e)6.922 F
1062: (xponential)-.165 E 4.38
1063: (annealing schedules permit resources to be spent adapti)72 615.392 R
1064: -.165(ve)-.275 G 4.38(ly on reannealing and on pacing the).165 F(con)72
1065: 637.392 Q -.165(ve)-.44 G -.198(rg).165 G 1.111(ence in all dimensions,\
1066: ensuring ample global searching in the \214rst phases of search and am\
1067: ple).198 F(quick con)72 659.392 Q -.165(ve)-.44 G -.198(rg).165 G
1068: (ence in the \214nal phases.).198 E 4.416(Ih)99.5 685.292 S -2.475 -.22
1069: (av e)-4.416 H 1.666(used ASA in se)4.636 F -.165(ve)-.275 G 1.666
1070: (ral systems, ranging from combat analysis \(Ingber).165 F 4.416(,F)-.44
1071: G 1.666(ujio, and W)-4.416 F(ehner)-.88 E(,)-.44 E .704
1072: (1991; Ingber and Sw)72 707.292 R(order)-.11 E 3.454(,1)-.44 G .704
1073: (991\), to \214nance \(Ingber)-3.454 F 3.454(,1)-.44 G .704(990; Ingber)
1074: -3.454 F 3.453(,W)-.44 G(ehner)-4.333 E F3 .703(et al)3.453 F F0 3.453
1075: (,1)C .703(991\), to neuroscience)-3.453 F(\(Ingber)72 729.292 Q 4.245
1076: (,1)-.44 G 1.496
1077: (991\), to a set of test problems \(Ingber and Rosen, 1992\), to a ne)
1078: -4.245 F 4.246(wt)-.275 G 1.496(echnique combining the)-4.246 F EP
1079: %%Page: 12 12
1080: %%BeginPageSetup
1081: BP
1082: %%EndPageSetup
1083: /F0 11/Times-Roman@0 SF(ASA Lessons Learned)72 58 Q 163.138(-12- Lester)
1084: 125.859 F(Ingber)2.75 E(po)72 94 Q(wer of SA with the ph)-.275 E
1085: (ysics of lar)-.055 E(ge-scale systems \(Ingber)-.198 E 2.75(,1)-.44 G
1086: (992\).)-2.75 E .041(An optimization algorithm typically is just one to\
1087: ol used in a major project, and man)99.5 119.9 R 2.79(yu)-.165 G .04
1088: (sers of ASA)-2.79 F 1.41(use this code as just one such tool.)72 141.9
1089: R 4.16(Furthermore, the)6.91 F 1.41(author has made it a polic)4.16 F
1090: 4.16(yn)-.165 G 1.41(ot to gi)-4.16 F 1.74 -.165(ve o)-.275 H 1.41
1091: (ut an).165 F(y)-.165 E .752(information he recei)72 163.9 R -.165(ve)
1092: -.275 G 3.501(sf).165 G .751(rom users unless the)-3.501 F 3.501(ys)
1093: -.165 G .751(peci\214cally permit him to do so.)-3.501 F .751(Thus, man)
1094: 6.251 F 3.501(yp)-.165 G .751(eople ask)-3.501 F 2.429(questions and gi)
1095: 72 185.9 R 2.76 -.165(ve f)-.275 H 2.43(eedback on the code the).165 F
1096: 5.18(ym)-.165 G 2.43(ight not gi)-5.18 F 2.76 -.165(ve o)-.275 H 5.18
1097: (therwise. In).165 F 2.43(order to get maximum)5.18 F 1.89(feedback wit\
1098: hout unduly bothering researchers, the e-mail ASA_list is not open foru\
1099: m, b)72 207.9 R 1.89(ut rather an)-.22 F(ef)72 229.9 Q .536
1100: (\214cient moderated medium to g)-.275 F .536(ather information.)-.055 F
1101: .536(Some published ackno)6.036 F .536(wledgments to use of the code)
1102: -.275 F .141(are in the asa_papers \214le of the ASA archi)72 251.9 R
1103: -.165(ve)-.275 G 2.891(;t).165 G .141(he disciplines range from ph)
1104: -2.891 F .141(ysics \(Bro)-.055 F(wn)-.275 E/F1 11/Times-Italic@0 SF .14
1105: (et al)2.891 F F0 2.89(,1)C .14(994; T)-2.89 F(ang)-.88 E F1 .328(et al)
1106: 72 273.9 R F0 3.078(,1)C .328(995\) to neural netw)-3.078 F .328
1107: (orks \(Cohen, 1994; Indi)-.11 F -.165(ve)-.275 G(ri).165 E F1 .328
1108: (et al)3.078 F F0 3.078(,1)C .329(993\) to dif)-3.078 F .329
1109: (\214cult imaging problems \(W)-.275 F 3.079(ua)-.55 G(nd)-3.079 E(Le)72
1110: 295.9 Q(vine, 1993\) to \214nance \(W)-.275 E(ofse)-.88 E 1.43 -.715
1111: (y, 1)-.165 H(993\), where most of the latter uses are k).715 E
1112: (ept proprietary)-.11 E(.)-.715 E/F2 11/Times-Bold@0 SF 2.75(3. ASA)72
1113: 339.9 R(OPTIONS)2.75 E 2.75(3.1. Options)72 387.8 R F0 1.412(ASA lik)
1114: 99.5 413.7 R 1.412(ely is the most po)-.11 F 1.411(werful and \215e)
1115: -.275 F 1.411(xible SA code presently a)-.165 F -.275(va)-.22 G 1.411
1116: (ilable, because the code has).275 F .298
1117: (bene\214ted from the feedback of the man)72 435.7 R 3.048(yu)-.165 G
1118: .299(sers, and their feedback has been used to add much to the code)
1119: -3.048 F(be)72 457.7 Q(yond the basic ASA algorithm described abo)-.165
1120: E -.165(ve)-.165 G(.).165 E .7(The code has tw)99.5 483.6 R 3.45(ob)-.11
1121: G .699(asic modules in the ASA C-code, a user and an asa module.)-3.45 F
1122: .699(All OPTIONS in)6.199 F .189(the code ha)72 505.6 R .519 -.165(ve b)
1123: -.22 H .189(een tested to w).165 F .189(ork with templates pro)-.11 F
1124: .189(vided in the user module.)-.165 F .19(Feedback has de)5.69 F -.165
1125: (ve)-.275 G .19(loped a).165 F .372
1126: (code which seems to run well across man)72 527.6 R 3.122(yp)-.165 G
1127: .372(latforms, e.g., PC')-3.122 F .372(s, Macs, Crays, man)-.605 F 3.122
1128: (yU)-.165 G .372(NIX w)-3.122 F(orkstations,)-.11 E(etc.)72 549.6 Q .408
1129: (The emphasis in de)99.5 575.5 R -.165(ve)-.275 G .408
1130: (lopment of ASA has been to add po).165 F .408(wer and \215e)-.275 F
1131: .408(xibility where)-.165 F -.165(ve)-.275 G 3.158(rp).165 G 3.159
1132: (ossible. T)-3.158 F(o)-.88 E(mak)72 597.5 Q 6.482(et)-.11 G 3.732
1133: (hese e)-6.482 F 3.731(xtra features and code accessible to non-e)-.165
1134: F 3.731(xpert programmers, a \231meta-language\232 of)-.165 F 1.245
1135: (OPTIONS is used.)72 619.5 R(Man)6.745 E 3.995(yo)-.165 G 3.995(ft)
1136: -3.995 G 1.245(hese OPTIONS can be set in the pro)-3.995 F 1.245
1137: (vided Mak)-.165 F 1.246(e\214le, an asa_opt data \214le)-.11 F .124
1138: (from which to read in information, ar)72 641.5 R .124
1139: (guments passed to the compilation procedures, or in the user module)
1140: -.198 F(\214les \(user)72 663.5 Q(.c, user)-.605 E(.h , and asa_user)
1141: -.605 E
1142: (.h, the latter being a bridge between the asa and user modules\).)-.605
1143: E 2.906(Ap)99.5 689.4 S .156(rice is paid for this continual de)-2.906 F
1144: -.165(ve)-.275 G .156(lopment of a more po).165 F .156
1145: (werful and more \215e)-.275 F .157(xible code.)-.165 F .157(The ne)
1146: 5.657 F(w)-.275 E .295(user is presented with man)72 711.4 R 3.045(yO)
1147: -.165 G .295(PTIONS, on the order of a hundred.)-3.045 F .295(In man)
1148: 5.795 F 3.045(yc)-.165 G .294(ases, when the ASA def)-3.045 F(ault)-.11
1149: E 1.811(OPTIONS w)72 733.4 R 1.812(ork \214ne, only the user')-.11 F
1150: 4.562(so)-.605 G 1.812(wn call to his/her cost function is required.)
1151: -4.837 F(Ho)7.312 E(we)-.275 E -.165(ve)-.275 G 2.692 -.44(r, i).165 H
1152: 4.562(ft).44 G(hese)-4.562 E EP
1153: %%Page: 13 13
1154: %%BeginPageSetup
1155: BP
1156: %%EndPageSetup
1157: /F0 11/Times-Roman@0 SF(ASA Lessons Learned)72 58 Q 163.138(-13- Lester)
1158: 125.859 F(Ingber)2.75 E(def)72 94 Q 2.306(aults are not suitable for a \
1159: particular system, then the user can become be)-.11 F 2.306
1160: (wildered by the man)-.275 F(y)-.165 E 3.048(OPTIONS. If)72 116 R .298
1161: (not much is kno)3.048 F(wn)-.275 E/F1 11/Times-Italic@0 SF 3.049(ap)
1162: 3.049 G(riori)-3.049 E<b4>-3.073 -2.343 M F0 .299
1163: (about the system to be optimized, then the task is to try to \214nd)
1164: 2.459 2.343 N .164(the v)72 138 R .164
1165: (alues of the OPTIONS appropriate to the gi)-.275 F -.165(ve)-.275 G
1166: 2.913(ns).165 G 2.913(ystem. The)-2.913 F .163(less kno)2.913 F .163
1167: (wn about the system, the harder)-.275 F(is this task.)72 160 Q 2.871(E\
1168: xperiences support the premise that the output of the code, using the A\
1169: SA_PRINT_MORE)99.5 185.9 R .293(OPTIONS to gi)72 207.9 R .622 -.165
1170: (ve i)-.275 H .292(nformation at each ne).165 F 3.042(wb)-.275 G .292
1171: (est accepted state, often can be used to diagnose problems in)-3.042 F
1172: 3.299(annealing. Ev)72 229.9 R(entually)-.165 E 3.299(,Ih)-.715 G .549
1173: (ope that enough e)-3.299 F .55
1174: (xperience will be generated, to be able to de)-.165 F -.165(ve)-.275 G
1175: .55(lop some kind).165 F 1.086(of graphical menu-dri)72 251.9 R -.165
1176: (ve)-.275 G 3.836(ne).165 G 1.086(xpert system to help guide users to o\
1177: ptimize a wide range of cost functions.)-4.001 F(Be)72 273.9 Q .514
1178: (yond this, I dream of a day when such a user)-.165 F .514
1179: (-friendly shell can guide users to alternati)-.22 F .845 -.165(ve o)
1180: -.275 H(ptimization).165 E(algorithms at dif)72 295.9 Q
1181: (ferent scales and stages of the optimization process.)-.275 E/F2 11
1182: /Times-Bold@0 SF 2.75(3.2. Examples)72 339.9 R(of OPTIONS)2.75 E F0
1183: 1.137(The follo)99.5 365.8 R 1.137
1184: (wing discussion of some of the OPTIONS a)-.275 F -.275(va)-.22 G 1.137
1185: (ilable in ASA also serv).275 F 1.137(es to illustrate the)-.165 F .356
1186: (typical kinds of problems man)72 387.8 R 3.106(yu)-.165 G .356(sers ha)
1187: -3.106 F .686 -.165(ve w)-.22 H .357
1188: (ith their particular systems, and some of the approaches that).165 F
1189: 5.414(SA can of)72 409.8 R 5.414(fer to f)-.275 F 5.414
1190: (ace these problems.)-.11 F 5.414(The OPTIONS are or)10.914 F -.055(ga)
1191: -.198 G 5.414(nized into three group.).055 F(The)10.913 E 3.212
1192: (DEFINE_OPTIONS comprise tw)72 431.8 R 5.962(os)-.11 G 3.213
1193: (et of OPTIONS, the Pre-Compile DEFINE_OPTIONS and the)-5.962 F 1.022(P\
1194: rinting DEFINE_OPTIONS, which are called at the time of compilation; th\
1195: ese comprise about half of)72 453.8 R .169(the OPTIONS.)72 475.8 R .169
1196: (The other Program OPTIONS are housed in a structure passed with the co\
1197: st function, and)5.669 F .49(together with the other parameters passed \
1198: in the cost function, these can be modi\214ed adapti)72 497.8 R -.165
1199: (ve)-.275 G(ly).165 E 5.989(.T)-.715 G .489(hat is,)-5.989 F(the)72
1200: 519.8 Q 2.75(yc)-.165 G(an be changed within the cost function to tak)
1201: -2.75 E 2.75(ee)-.11 G -.275(ff)-2.75 G
1202: (ect upon reentering the asa program.).275 E F2 2.75(3.2.1. Integer)72
1203: 563.8 R(and continuous parameters)2.75 E F0 .315
1204: (ASA can accommodate mixtures of inte)99.5 589.7 R .316
1205: (ger and continuous parameters.)-.165 F .316(This is accomplished quite)
1206: 5.816 F(simply)72 611.7 Q 2.909(,w)-.715 G .159(ith a small o)-2.909 F
1207: -.165(ve)-.165 G .159(rhead for inte).165 F .158
1208: (gers, by truncating generated \215oating-point numbers within sensible)
1209: -.165 F(inte)72 633.7 Q 2.612(gral windo)-.165 F 5.362(ws. There)-.275 F
1210: (ha)5.362 E 2.942 -.165(ve b)-.22 H 2.612(een man).165 F 5.362<7999>
1211: -.165 G 2.612(rumors\232 that SA can only handle inte)-5.362 F 2.612
1212: (ger or continuous)-.165 F(parameters, b)72 655.7 Q
1213: (ut these statements are unsupported.)-.22 E EP
1214: %%Page: 14 14
1215: %%BeginPageSetup
1216: BP
1217: %%EndPageSetup
1218: /F0 11/Times-Roman@0 SF(ASA Lessons Learned)72 58 Q 163.138(-14- Lester)
1219: 125.859 F(Ingber)2.75 E/F1 11/Times-Bold@0 SF 2.75(3.2.2. Constraints)72
1220: 94 R F0 1.037(One of the immediate attractions of SA to people trying t\
1221: o optimize comple)99.5 119.9 R 3.787(xs)-.165 G 1.037
1222: (ystems is the ease)-3.787 F .603(with which SA can accommodate comple)
1223: 72 141.9 R 3.353(xc)-.165 G 3.353(onstraints. T)-3.353 F(ypically)-.88 E
1224: 3.353(,t)-.715 G .603(here is no need for penalty functions,)-3.353 F
1225: 3.311(etc. Generated)72 163.9 R .561(points that do not satisfy the con\
1226: straints are simply rejected before trying an)3.311 F 3.31(ya)-.165 G
1227: (cceptance)-3.31 E(test.)72 185.9 Q .075
1228: (Equality constraints, if processed as abo)99.5 211.8 R -.165(ve)-.165 G
1229: 2.826(,p).165 G .076(resent a problem for an)-2.826 F 2.826(yg)-.165 G
1230: .076(lobal optimization that relies)-2.826 F 2.044
1231: (on sampling, because the search is being constrained on the surf)72
1232: 233.8 R 2.043(ace of some v)-.11 F 2.043(olume, and the entire)-.22 F
1233: -.22(vo)72 255.8 S 3.349(lume is being sampled.).22 F 3.35(Therefore, i\
1234: t is recommended that the user \214rst numerically substitute)8.849 F
1235: 3.165(solution\(s\) of the the equalities for some parameters.)72 277.8
1236: R -.165(Fo)8.665 G 5.915(re).165 G 3.165(xample, if the cost function)
1237: -6.08 F/F2 11/Times-Italic@0 SF(C)5.739 E F0(has)6.706 E F2(n)6.31 E F0
1238: (parameters,)72 299.8 Q F2(C)4.054 E F0(\().792 E F2(p)1.375 E/F3 8
1239: /Times-Roman@0 SF(1).209 2.2 M F0(,).55 -2.2 M F2(p)3.245 E F3(2).209
1240: 2.2 M F0(,).55 -2.2 M 1.87(...)1.87 -2.371 O(,)2.371 I F2(p)3.245 E/F4 8
1241: /Times-Italic@0 SF(n).497 2.2 M F0 1.481
1242: (\), and an equality constraint e).742 -2.2 N 1.481
1243: (xists between parameters)-.165 F F2(p)5.606 E F4(n).497 2.2 M F0(and)
1244: 4.973 -2.2 M F2(p)5.606 E F4(n).497 2.2 M/F5 8/Symbol SF(-).192 E F3(1)A
1245: F0 4.231(,t).55 -2.2 O(hen)-4.231 E(solv)72 321.8 Q 4.263(et)-.165 G
1246: 1.513(his equation for)-4.263 F F2(p)5.638 E F4(n).497 2.2 M F0 4.263
1247: (,n).742 -2.2 O 1.513(umerically or algebraically)-4.263 F 4.263(,r)
1248: -.715 G 1.513(ede\214ning the cost function to one with)-4.263 F F2(n)
1249: 4.659 E/F6 11/Symbol SF(-)2.684 E F0(1)2.42 E(parameters,)72 343.8 Q F2
1250: (C)3.158 E F6<a2>.792 E F0 6.084(.I)C 3.334(ft)-6.084 G .584
1251: (he solution to this equation, or perhaps a set of)-3.334 F F2(m)3.753 E
1252: F0 .585(such equality constraints to reduce)3.687 F 1.422
1253: (the number of parameters actually processed by ASA to)72 365.8 R F2(n)
1254: 4.567 E F6(-)2.684 E F2(m)2.838 E F0 4.171(,i).352 G 4.171(sn)-4.171 G
1255: 1.421(ot simply written do)-4.171 F 1.421(wn, then such)-.275 F
1256: (constraints must be solv)72 387.8 Q
1257: (ed with other algorithms within the cost function.)-.165 E F1 2.75
1258: (3.2.3. Annealing)72 431.8 R(scales)2.75 E F0 .572(Perhaps the easiest \
1259: to understand problem that can arise when using SA, also is the the mos\
1260: t often)99.5 457.7 R(ne)72 479.7 Q 4.027(glected. I)-.165 F 1.277
1261: (get man)4.027 F 4.027(yq)-.165 G 1.276
1262: (ueries that start out questioning wh)-4.027 F 4.026(yA)-.055 G 1.276
1263: (SA doesn')-4.026 F 4.026(ti)-.198 G 1.276(mmediately \214nd the global)
1264: -4.026 F 1.667(optimal point?)72 501.7 R 1.668(The answer most often li\
1265: es in the scaling parameters used in annealing the parameter)7.168 F
1266: (and/or cost temperatures.)72 523.7 Q -.165(Fo)99.5 549.6 S 3.148(re)
1267: .165 G .398(xample, if the search is carried out in a system with se)
1268: -3.313 F -.165(ve)-.275 G .397(ral local minima, b).165 F .397
1269: (ut the temperature)-.22 F .039(is too lo)72 571.6 R 2.789(ws)-.275 G
1270: 2.79(ot)-2.789 G .04
1271: (hat only rarely can the search sample these minima, it may tak)-2.79 F
1272: 2.79(ea)-.11 G 2.79(ne)-2.79 G .04(xtremely long time with)-2.955 F .76
1273: (arbitrarily good numerical precision to e)72 593.6 R -.165(ve)-.275 G
1274: .759(ntually sample these minima as normal annealing proceeds to).165 F
1275: (lo)72 615.6 Q .605(wer and lo)-.275 F .606(wer temperatures.)-.275 F
1276: (Clearly)6.106 E 3.356(,i)-.715 G 3.356(tw)-3.356 G .606
1277: (ould be better to ha)-3.466 F .936 -.165(ve t)-.22 H .606
1278: (he starting temperature at the scale in).165 F
1279: (question be commensurately lar)72 637.6 Q(ger)-.198 E 2.75(,a)-.44 G
1280: (nd perhaps be cooled more slo)-2.75 E(wly)-.275 E(.)-.715 E(\(a\))99.5
1281: 663.5 Q F2 -.88(Pa)3.167 G -.165(ra).88 G .417(meter temper).165 F(atur)
1282: -.165 E(es)-.407 E F0 5.917(.I)C 3.167(ns)-5.917 G .417
1283: (ome SA algorithms, lik)-3.167 F 3.167(eB)-.11 G .416
1284: (A, the starting temperature controls the)-3.552 F -.275(va)72 685.5 S
1285: .151(lues of temperature encountered at subsequent stages of search.)
1286: .275 F .152(In ASA, because of the \214nite ranges of)5.652 F .866(the \
1287: parameters, the parameter temperatures are started to establish to a f)
1288: 72 707.5 R .866(at tail throughout the range; the)-.11 F -.165(ex)72
1289: 729.5 S .623(ponential annealing rates usually permit selecting e).165 F
1290: -.165(ve)-.275 G 3.374(nq).165 G .624(uite lar)-3.374 F .624
1291: (ge initial ranges to be sure of co)-.198 F -.165(ve)-.165 G(ring).165 E
1292: EP
1293: %%Page: 15 15
1294: %%BeginPageSetup
1295: BP
1296: %%EndPageSetup
1297: /F0 11/Times-Roman@0 SF(ASA Lessons Learned)72 58 Q 163.138(-15- Lester)
1298: 125.859 F(Ingber)2.75 E .036(all optima.)72 94 R .036
1299: (There are free ASA-parameters for each temperature to scale its e)5.536
1300: F .035(xponential decrease, without)-.165 F(af)72 116 Q
1301: (fecting the basic sampling proof.)-.275 E(\(b\))99.5 141.9 Q/F1 11
1302: /Times-Italic@0 SF 2.018(Cost temper)4.768 F(atur)-.165 E(e)-.407 E F0
1303: 7.518(.T)C 2.019(he annealing scale for the cost temperature, also call\
1304: ed the acceptance)-7.518 F 1.338(temperature, af)72 163.9 R 1.338
1305: (fects the rate of narro)-.275 F 1.338(wing the windo)-.275 F 4.088(wo)
1306: -.275 G 4.088(ft)-4.088 G 1.338(he Boltzmann acceptance test.)-4.088 F
1307: 1.338(In ASA, this)6.838 F(scale can be adapti)72 185.9 Q -.165(ve)-.275
1308: G(ly changed, and e).165 E -.165(ve)-.275 G 2.75(nt).165 G
1309: (he Boltzmann test can be changed to a dif)-2.75 E(ferent distrib)-.275
1310: E(ution.)-.22 E/F2 11/Times-Bold@0 SF 2.75(3.2.4. Reannealing)72 229.9 R
1311: F0(\(a\))99.5 255.8 Q F1 -.88(Pa)3.33 G -.165(ra).88 G .58(meter temper)
1312: .165 F(atur)-.165 E(es)-.407 E F0 6.08(.F)C .581
1313: (or the parameter temperatures, the tangents \(or an)-6.245 F 3.331(yo)
1314: -.165 G .581(ther alternati)-3.331 F -.165(ve)-.275 G 1.451
1315: (functions that might be de\214ned by the user\) are used as a relati)72
1316: 277.8 R 1.78 -.165(ve m)-.275 H 1.45
1317: (easure of the \231steepness\232 of each).165 F 2.598
1318: (dimension the most recent best sa)72 299.8 R -.165(ve)-.22 G 5.348(ds)
1319: .165 G 5.348(tate. As)-5.348 F 2.598
1320: (demonstrated for the ASA_TEST problem \(Ingber)5.348 F(,)-.44 E
1321: (1993b\), this feature can enhance the ef)72 321.8 Q(\214cienc)-.275 E
1322: 2.75(yo)-.165 G 2.75(ft)-2.75 G(he search.)-2.75 E(\(b\))99.5 347.7 Q F1
1323: .596(Cost temper)3.346 F(atur)-.165 E(e)-.407 E F0 6.096(.F)C .596(or t\
1324: he cost temperature, a separate OPTIONS permits rescaling of the cost)
1325: -6.261 F .421(temperature to be set to the scale of the minimum of the \
1326: current cost temperature and the absolute v)72 369.7 R(alues)-.275 E
1327: .339(of the last and best sa)72 391.7 R -.165(ve)-.22 G 3.089(dm).165 G
1328: .339(inima, to k)-3.089 F .338(eep the acceptance test sensiti)-.11 F
1329: .668 -.165(ve a)-.275 H 3.088(tar).165 G .338(easonable scale.)-3.088 F
1330: .338(This can be)5.838 F -.165(ex)72 413.7 S .085
1331: (tremely important of the system').165 F 2.835(st)-.605 G .086
1332: (errain changes with the scale of the search.)-2.835 F .086
1333: (This procedure also may)5.586 F .055(need to radically altered, possib\
1334: le with other OPTIONS, if the search early becomes stuck in local optim\
1335: a,)72 435.7 R(e.g., because the system')72 457.7 Q 2.75(st)-.605 G
1336: (errain abruptly changes with the scale of the search.)-2.75 E F2 2.75
1337: (3.2.5. Quenching)72 501.7 R F0 .106(An SA algorithm loses much of its \
1338: authority if the search \231cheats\232 by trying to anneal at rates f)
1339: 99.5 527.6 R(aster)-.11 E .738(than permitted by its associated proof, \
1340: e.g., simulated quenching \(SQ\).)72 549.6 R(Ho)6.238 E(we)-.275 E -.165
1341: (ve)-.275 G 1.618 -.44(r, t).165 H .738(his can be useful in a).44 F
1342: (number of circumstances \(Ingber)72 571.6 Q 2.75(,1)-.44 G(993b\).)
1343: -2.75 E .324(When the dimension of a parameter space, each parameter ha)
1344: 99.5 597.5 R .325(ving a continuous or lar)-.22 F .325(ge inte)-.198 F
1345: .325(gral set)-.165 F .64(of v)72 619.5 R .64
1346: (alues, reaches 15 to 20, the v)-.275 F .639
1347: (olume of search typically becomes quite lar)-.22 F .639
1348: (ge and this can se)-.198 F -.165(ve)-.275 G .639(rely tax).165 F 1.283
1349: (most present-day w)72 641.5 R 4.033(orkstations. Instead)-.11 F 1.283
1350: (of just gi)4.033 F 1.284(ving up on SA and trying a dif)-.275 F 1.284
1351: (ferent \231greedy\232 and/or)-.275 F(quasi-Ne)72 663.5 Q
1352: (wton algorithm, ASA pro)-.275 E(vides a methodical w)-.165 E(ay to de)
1353: -.11 E(viate from SA into SQ algorithms.)-.275 E .962
1354: (As another use of quenching, one that does not necessarily violate an)
1355: 99.5 689.4 R 3.712(ys)-.165 G .962(ampling proof, it may be)-3.712 F .39
1356: (useful in the course of search to adapti)72 711.4 R -.165(ve)-.275 G
1357: .39(ly drop subsets of parameters that seem to ha).165 F .72 -.165(ve b)
1358: -.22 H .39(een reasonably).165 F .832(optimized relati)72 733.4 R 1.162
1359: -.165(ve t)-.275 H 3.582(oo).165 G .832(ther parameters.)-3.582 F .831
1360: (The remaining parameters can then be more ef)6.331 F .831
1361: (\214ciently searched)-.275 F EP
1362: %%Page: 16 16
1363: %%BeginPageSetup
1364: BP
1365: %%EndPageSetup
1366: /F0 11/Times-Roman@0 SF(ASA Lessons Learned)72 58 Q 163.138(-16- Lester)
1367: 125.859 F(Ingber)2.75 E 3.268(within their smaller dimensional space, b\
1368: y adjusting the dependence of the annealing to the ne)72 94 R(w)-.275 E
1369: 2.75(dimension. This)72 116 R(can be accomplished con)2.75 E -.165(ve)
1370: -.44 G(niently with the Q).165 E(UENCHing OPTIONS.)-.11 E/F1 11
1371: /Times-Bold@0 SF 2.75(3.2.6. ASA)72 160 R(sampling)2.75 E F0 .166(Since\
1372: ASA accomplishes its \214t by importance sampling the space of paramet\
1373: ers, it w)99.5 185.9 R .166(ould seem that)-.11 F .544
1374: (this process should pro)72 207.9 R .544(vide a good sampling technique\
1375: for other purposes, e.g., performing inte)-.165 F 3.295(grals. Of)-.165
1376: F 1.539(course, as stated abo)72 229.9 R -.165(ve)-.165 G 4.288(,t).165
1377: G 1.538(he use of Monte Carlo techniques for performing inte)-4.288 F
1378: 1.538(grals \(Metropolis)-.165 F/F2 11/Times-Italic@0 SF 1.538(et al)
1379: 4.288 F F0(,)A .254
1380: (1953\) is generally credited to be the origin of the de)72 251.9 R
1381: -.165(ve)-.275 G .254(lopment of SA \(Kirkpatrick).165 F F2 .254(et al)
1382: 3.004 F F0 3.004(,1)C 3.004(983\). Ho)-3.004 F(we)-.275 E -.165(ve)-.275
1383: G -.44(r,).165 G 2.85(importance sampling with the f)72 273.9 R 2.849
1384: (astest permitted temperature schedules often can lead to quite poor)
1385: -.11 F .568(resolutions of local minima which may substantially contrib)
1386: 72 295.9 R .568(ute to inte)-.22 F 3.319(grals. Then,)-.165 F .569
1387: (the rates of annealing)3.319 F 1.267(must be slo)72 317.9 R 1.267
1388: (wed do)-.275 F 1.266(wn, e.g., using in)-.275 F -.165(ve)-.44 G 1.266
1389: (rse Q).165 F 1.266(UENCHing, to get better resolution.)-.11 F 1.266
1390: (The ASA__SAMPLE)6.766 F .092(OPTIONS collects the generating and accep\
1391: tance biases incurred during importance sampling, so that this)72 339.9
1392: R(information can be used more generally than for just \214nding the op\
1393: timal point of the \214t.)72 361.9 Q F1 2.75(3.2.7. Self)72 405.9 R
1394: (optimization)2.75 E F0 .25(An adv)99.5 431.8 R .25(antage of C code o)
1395: -.275 F -.165(ve)-.165 G 3(rs).165 G .249
1396: (ome other languages is the relati)-3 F .579 -.165(ve e)-.275 H .249
1397: (ase by which recursi).165 F .579 -.165(ve c)-.275 H .249(alls can).165
1398: F .369(be implemented.)72 453.8 R .369(Some care must be tak)5.869 F
1399: .369(en to k)-.11 F .369(eep v)-.11 F .369
1400: (ariables local to each subroutine.)-.275 F .37(In its current form)
1401: 5.869 F .442(ASA can recursi)72 475.8 R -.165(ve)-.275 G .442
1402: (ly call itself.).165 F .441(Some comple)5.941 F 3.191(xp)-.165 G .441
1403: (roblems, possessing nests of optimized systems, require)-3.191 F(this.)
1404: 72 497.8 Q 1.24(If not much information is kno)99.5 523.7 R 1.241
1405: (wn about a particular system, if the ASA def)-.275 F 1.241
1406: (aults do not seem to)-.11 F -.11(wo)72 545.7 S .002(rk v).11 F .002
1407: (ery well, and if after a bit of e)-.165 F .002
1408: (xperimentation it still is not clear ho)-.165 F 2.752(wt)-.275 G 2.751
1409: (os)-2.752 G .001(elect v)-2.751 F .001(alues for some of the)-.275 F
1410: 1.991(ASA OPTIONS, then the SELF_OPTIMIZE OPTIONS can be v)72 567.7 R
1411: 1.991(ery useful.)-.165 F 1.992(This sets up a top le)7.491 F -.165(ve)
1412: -.275 G(l).165 E .778(search on the ASA OPTIONS themselv)72 589.7 R .777
1413: (es, using criteria of the system as its o)-.165 F .777
1414: (wn cost function, e.g., the)-.275 F 1.001(best attained optimal v)72
1415: 611.7 R 1.001(alue of the system')-.275 F 3.751(sc)-.605 G 1.001
1416: (ost function \(the cost function for the actual problem to be)-3.751 F
1417: (solv)72 633.7 Q .958(ed\) for each gi)-.165 F -.165(ve)-.275 G 3.708
1418: (ns).165 G .958(et of top le)-3.708 F -.165(ve)-.275 G 3.708(lO).165 G
1419: .958(PTIONS, or the number of generated states required to reach a)
1420: -3.708 F(gi)72 655.7 Q -.165(ve)-.275 G 4.721(nv).165 G 1.971
1421: (alue of the system')-4.996 F 4.721(sc)-.605 G 1.971(ost function, etc.)
1422: -4.721 F 1.971(Since this can consume a lot of CPU resources, it is)
1423: 7.471 F .444(recommended that only a fe)72 677.7 R 3.194(wA)-.275 G .444
1424: (SA OPTIONS and a scaled do)-3.194 F .443
1425: (wn system cost function or system data be)-.275 F
1426: (selected for this OPTIONS.)72 699.7 Q EP
1427: %%Page: 17 17
1428: %%BeginPageSetup
1429: BP
1430: %%EndPageSetup
1431: /F0 11/Times-Roman@0 SF(ASA Lessons Learned)72 58 Q 163.138(-17- Lester)
1432: 125.859 F(Ingber)2.75 E(Ev)99.5 94 Q 1.668(en if good results are being\
1433: attained by ASA, SELF_OPTIMIZE can be used to \214nd a more)-.165 F(ef)
1434: 72 116 Q .832(\214cient set of ASA OPTIONS.)-.275 F 3.582(It)6.332 G
1435: .832(hink that this kind of OPTIONS w)-3.582 F .832
1436: (ould be useful for most useful for)-.11 F(man)72 138 Q 3.751(yn)-.165 G
1437: 1.001(on-linear optimization algorithms.)-3.751 F(Man)6.501 E 3.751(yo)
1438: -.165 G 3.751(ft)-3.751 G 1.001(he OPTIONS brok)-3.751 F 1.001
1439: (en out in clear vie)-.11 F 3.752(wi)-.275 G 3.752(nA)-3.752 G 1.002
1440: (SA are)-3.752 F 2.833(similarly represented b)72 160 R 2.833
1441: (ut \231hidden\232 within the code)-.22 F 2.832(of other algorithms.)
1442: 8.332 F 2.832(Self optimization of such)8.332 F(parameters can be v)72
1443: 182 Q(ery useful for production runs of comple)-.165 E 2.75(xs)-.165 G
1444: (ystems.)-2.75 E/F1 11/Times-Bold@0 SF 2.75(3.2.8. Alter)72 226 R(nati)
1445: -.165 E .22 -.11(ve d)-.11 H(istrib).11 E(utions/functions)-.22 E F0
1446: 1.146(There are OPTIONS to permit replacing or modifying the functions \
1447: and distrib)99.5 251.9 R 1.146(utions used in the)-.22 F 1.566
1448: (asa module.)72 273.9 R -.165(Fo)7.066 G 4.316(ee).165 G 1.566(xample, \
1449: modi\214cations can be made of the generating function \(e.g., v)-4.481
1450: F 1.566(ariants of the)-.275 F .677(Boltzmann and Cauch)72 295.9 R 3.427
1451: (yd)-.055 G(istrib)-3.427 E .677(utions are gi)-.22 F -.165(ve)-.275 G
1452: 3.427(ni).165 G 3.427(nt)-3.427 G .677
1453: (he user module\), the acceptance function \(e.g., a class)-3.427 F
1454: 1.084
1455: (of functions that asymptotically approach the Boltzmann function is gi)
1456: 72 317.9 R -.165(ve)-.275 G 3.834(ni).165 G 3.834(nt)-3.834 G 1.084
1457: (he user module\), and the)-3.834 F(reannealing functions used to resca\
1458: le the parameter and cost temperatures.)72 339.9 Q F1 2.75(3.2.9. P)72
1459: 383.9 R(arallel code)-.11 E F0 1.072(It is quite dif)99.5 409.8 R 1.072
1460: (\214cult to directly parallelize an SA algorithm \(Ingber)-.275 F 3.822
1461: (,1)-.44 G 1.072(993b\), e.g., without incurring)-3.822 F -.165(ve)72
1462: 431.8 S .607(ry restricti).165 F .937 -.165(ve c)-.275 H .606
1463: (onstraints on temperature schedules \(Kimura and T).165 F .606
1464: (aki, 1991\), or violating an associated)-.88 F 3.253
1465: (sampling proof \(Frost, 1993\).)72 453.8 R(Ho)8.753 E(we)-.275 E -.165
1466: (ve)-.275 G 4.133 -.44(r, t).165 H 3.253(he f).44 F 3.253
1467: (at tail of ASA permits parallelization of de)-.11 F -.165(ve)-.275 G
1468: (loping).165 E 1.285(generated states prior to subjecting them to the a\
1469: cceptance test \(Ingber)72 475.8 R 4.035(,1)-.44 G 4.034(992\). The)
1470: -4.035 F(ASA_P)4.034 E(ARALLEL)-1.012 E 1.05(OPTIONS pro)72 497.8 R
1471: 1.051(vide parameters to easily parallelize the code, using v)-.165 F
1472: 1.051(arious implementations, e.g., PVM,)-.275 F(shared memory)72 519.8
1473: Q 2.75(,e)-.715 G(tc.)-2.75 E 1.205(The scale of parallelization af)99.5
1474: 545.7 R 1.204
1475: (forded by ASA, without violating its sampling proof, is gi)-.275 F
1476: -.165(ve)-.275 G 3.954(nb).165 G 3.954(ya)-3.954 G .426
1477: (typical ratio of the number of generated to accepted states.)72 567.7 R
1478: (Se)5.927 E -.165(ve)-.275 G .427(ral e).165 F .427
1479: (xperts in parallelization suggest that)-.165 F(massi)72 589.7 Q .655
1480: -.165(ve p)-.275 H .324
1481: (arallelization e.g., on the order of the human brain, may tak).165 F
1482: 3.074(ep)-.11 G .324(lace quite f)-3.074 F .324
1483: (ar into the future, that)-.11 F .978(this might be some)72 611.7 R .978
1484: (what less useful for man)-.275 F 3.729(ya)-.165 G .979
1485: (pplications than pre)-3.729 F .979
1486: (viously thought, and that most useful)-.275 F .25
1487: (scales of parallelization might be on scales of order 10 to 1000.)72
1488: 633.7 R .249(Depending on the speci\214c problem, such)5.749 F(scales a\
1489: re common in ASA optimization, and the current ASA code can implement s\
1490: uch parallelization.)72 655.7 Q .539(No speci\214c parallel implementat\
1491: ion has yet been included in the code.)99.5 681.6 R 3.29(Ap)6.039 G .54
1492: (roject w)-3.29 F .54(as set up under)-.11 F .177(an National Science F)
1493: 72 703.6 R .177(oundation \(NSF\) grant of Cray time at the Pittsb)-.165
1494: F(ur)-.22 E .177(gh Supercomputer Center for this)-.198 F .645
1495: (purpose, b)72 725.6 R .646
1496: (ut we soon realized that the grant rules restricted us to \214rst v)
1497: -.22 F .646(ectorize the ASA code on a C90.)-.165 F EP
1498: %%Page: 18 18
1499: %%BeginPageSetup
1500: BP
1501: %%EndPageSetup
1502: /F0 11/Times-Roman@0 SF(ASA Lessons Learned)72 58 Q 163.138(-18- Lester)
1503: 125.859 F(Ingber)2.75 E .273(This seemed lik)72 94 R 3.023(eal)-.11 G
1504: .273(ot of ef)-3.023 F .272
1505: (fort for a speci\214c architecture, e.g., an ef)-.275 F .272
1506: (fort that w)-.275 F .272(ould not e)-.11 F -.165(ve)-.275 G 3.022(nt)
1507: .165 G .272(ranslate into)-3.022 F 1.904
1508: (useful code on a T3D, and yet this still w)72 116 R 1.905
1509: (ould not address the true parallelization of the ASA code.)-.11 F .421
1510: (Therefore, our v)72 138 R .421(olunteer group went as f)-.22 F .42
1511: (ar as to con\214rm the nature of the required code, and I placed the)
1512: -.11 F(abo)72 160 Q .433 -.165(ve \231)-.165 H .103
1513: (hooks\232 into the present code.).165 F .104
1514: (People who played a major role in e)5.604 F .104
1515: (xamining these issues were T)-.165 F(im)-.385 E .045
1516: (Burns of the Utah Supercomputing Institute, Alan Cabrera of Sanw)72 182
1517: R 2.795(aF)-.11 G .045(inancial Products Co., and W)-2.795 F(olfram)-.88
1518: E(Gloger of Poliklinik fu)72 204 Q -1.836 .918(\250r Z)-4.581 H
1519: (ahnerhaltung und P)-.918 E(arodontologie.)-.165 E/F1 11/Times-Bold@0 SF
1520: 2.75(4. IV)72 248 R 2.75(.M)-1.595 G(ISLEADING SA CLAIMS)-2.75 E 2.75
1521: (4.1. General)72 295.9 R(Considerations)2.75 E 2.75(4.1.1. Comparisons)
1522: 72 343.8 R(among algorithms)2.75 E F0 1.773
1523: (There is not much doubt that, for an)99.5 369.7 R 4.523(yr)-.165 G
1524: 1.773(easonably dif)-4.523 F 1.773
1525: (\214cult nonlinear or stochastic system, some)-.275 F .324
1526: (optimization codes will perform better than others.)72 391.7 R .323
1527: (Nonlinear systems are typically non-typical, and so it)5.823 F .354
1528: (should be e)72 413.7 R .354(xpected that some algorithms are better su\
1529: ited for some systems than for others.)-.165 F .355(It is unlik)5.855 F
1530: (ely)-.11 E(that there e)72 435.7 Q -.165(ve)-.275 G 2.75(rw).165 G
1531: (ill be de)-2.75 E -.165(ve)-.275 G
1532: (loped a \231black-box canned\232 algorithm as e).165 E(xist for man)
1533: -.165 E 2.75(yl)-.165 G(inear systems.)-2.75 E .368
1534: (Thus, comparisons among algorithms on simple \231to)99.5 461.6 R .368
1535: (y\232 problems, while the)-.11 F 3.117(ym)-.165 G .367
1536: (ay be interesting and)-3.117 F 1.799
1537: (necessary \(and therefore are included belo)72 483.6 R 1.8
1538: (w\), perhaps e)-.275 F 1.8
1539: (xposing something of the nature of the tested)-.165 F .489
1540: (algorithms, may of)72 505.6 R .488(fer little insight to help a gi)
1541: -.275 F -.165(ve)-.275 G 3.238(nr).165 G .488(esearcher f)-3.238 F .488
1542: (aced with a ne)-.11 F 3.238(wp)-.275 G 3.238(roblem. Most)-3.238 F(lik)
1543: 3.238 E(ely)-.11 E 3.238(,t)-.715 G(he)-3.238 E(researcher must w)72
1544: 527.6 Q(ork to learn more about the system to best apply a gi)-.11 E
1545: -.165(ve)-.275 G 2.75(na).165 G(lgorithm.)-2.75 E .264
1546: (There is a formal proof that all algorithms should perform the same o)
1547: 99.5 553.5 R -.165(ve)-.165 G 3.015(ra).165 G .265(ll problems \(W)
1548: -3.015 F .265(olpert and)-.88 F(Macready)72 575.5 Q 4.893(,1)-.715 G
1549: 4.893(995\). This)-4.893 F(ob)4.893 E 2.142
1550: (viously is of little practical help for a gi)-.165 F -.165(ve)-.275 G
1551: 4.892(np).165 G 2.142(articular problem, albeit the)-4.892 F .085
1552: (authors claim that their approach of)72 597.5 R .085
1553: (fers some guidelines for speci\214c algorithms.)-.275 F -.165(Fo)5.586
1554: G 2.836(re).165 G .086(xample, it w)-3.001 F .086(ould be)-.11 F .229
1555: (ill-advised to dra)72 619.5 R 2.979(wt)-.165 G .229
1556: (he conclusion that therefore one might as well al)-2.979 F -.11(wa)-.11
1557: G .228(ys just use Ne).11 F(wton')-.275 E 2.978(sa)-.605 G .228
1558: (lgorithm on)-2.978 F .383(all problems f)72 641.5 R .383
1559: (aced by a research group.)-.11 F(Ho)5.883 E(we)-.275 E -.165(ve)-.275 G
1560: 1.263 -.44(r, s).165 H .383
1561: (uch research is necessary to bring more objecti).44 F .384(vity to)
1562: -.275 F(the present \231art\232 of global optimization.)72 663.5 Q EP
1563: %%Page: 19 19
1564: %%BeginPageSetup
1565: BP
1566: %%EndPageSetup
1567: /F0 11/Times-Roman@0 SF(ASA Lessons Learned)72 58 Q 163.138(-19- Lester)
1568: 125.859 F(Ingber)2.75 E/F1 11/Times-Bold@0 SF 2.75(4.1.2. T)72 94 R
1569: (uning algorithms f)-1.012 E(or speci\214c pr)-.275 E(oblems)-.198 E F0
1570: 1.094(There is not much doubt that, for an)99.5 119.9 R 3.844(yr)-.165 G
1571: 1.094(easonably dif)-3.844 F 1.093
1572: (\214cult nonlinear or stochastic system, a gi)-.275 F -.165(ve)-.275 G
1573: (n).165 E .657(optimization algorithm can be \231tuned\232 to enhance i\
1574: ts performance.)72 141.9 R .658(Indeed, since it tak)6.157 F .658
1575: (es some time and)-.11 F(ef)72 163.9 Q .26(fort to become f)-.275 F .26
1576: (amiliar with a gi)-.11 F -.165(ve)-.275 G 3.01(nc).165 G .26
1577: (ode, the ability to tune a gi)-3.01 F -.165(ve)-.275 G 3.01(na).165 G
1578: .26(lgorithm for use in more than one)-3.01 F
1579: (problem should be considered an important feature of an algorithm.)72
1580: 185.9 Q F1 2.75(4.2. Examples)72 229.9 R(of Comparisons)2.75 E 2.75
1581: (4.2.1. Genetic)72 277.8 R(algorithms \(GA\))2.75 E F0 4.629(Ad)99.5
1582: 303.7 S 1.879(irect comparison w)-4.629 F 1.88
1583: (as made between ASA/VFSR and a publicly a)-.11 F -.275(va)-.22 G 1.88
1584: (ilable genetic algorithm).275 F .611(\(GA\) code, using a test suite a\
1585: lready adapted and adopted for GA \(Schraudolph and Grefenstette, 1991\
1586: \).)72 325.7 R(In each case, ASA outperformed the GA problem.)72 347.7 Q
1587: .805(It should be clear that GA is a class of algorithms that are inter\
1588: esting in their o)99.5 373.6 R .806(wn right.)-.275 F .806(GA w)6.306 F
1589: (as)-.11 E .524(not originally de)72 395.6 R -.165(ve)-.275 G .524
1590: (loped as an optimization algorithm \(De).165 F .524
1591: (Jong, 1992\), and basic GA does not of)5.5 F .523(fer an)-.275 F(y)
1592: -.165 E .359(statistical guarantee of global con)72 417.6 R -.165(ve)
1593: -.44 G -.198(rg).165 G .359(ence to an optimal point \(F).198 F .36
1594: (orrest, 1993\).)-.165 F(Ne)5.86 E -.165(ve)-.275 G .36
1595: (rtheless, it should be).165 F -.165(ex)72 439.6 S
1596: (pected that GA may be better suited for some problems than SA.).165 E
1597: F1 2.75(4.2.2. Comparing)72 483.6 R -.33(BA)2.75 G 2.75(,F).33 G
1598: (A, and ASA)-3.74 E F0 3.944(As)99.5 509.5 S 1.194(tudy w)-3.944 F 1.194
1599: (as made comparing the performance among Boltzmann annealing \(B)-.11 F
1600: 1.194(A\), f)-.385 F 1.193(ast annealing)-.11 F(\(F)72 531.5 Q .239
1601: (A\), and ASA, using the dif)-.814 F .239
1602: (\214cult test problem that comes with the ASA code \(Rosen, 1992\).)
1603: -.275 F .24(Only ASA)5.739 F(re)72 553.5 Q 1.918
1604: (gularly attained the global minimum, and it w)-.165 F 1.918(as more ef)
1605: -.11 F 1.918(\214cient in attaining re)-.275 F 1.918
1606: (gular minima at each)-.165 F 1.523
1607: (comparable number of generated states than B)72 576.178 R 4.274(Ao)
1608: -.385 G 4.274(rF)-4.274 G 4.274(A. The)-5.088 F 1.524
1609: (ASA_TEST problem contains 10)4.274 F/F2 8/Times-Roman@0 SF(20)-4.62 I
1610: F0(local)4.824 4.62 M .151
1611: (minima with a parameter dimension of 4 \(Corana)72 598.178 R/F3 11
1612: /Times-Italic@0 SF .151(et al)2.901 F F0 2.901(,1)C .151
1613: (987\), and it typically tak)-2.901 F .15(es 2000\2553000 generated)-.11
1614: F(states, or about 2 CPU-sec on a Sun SP)72 620.178 Q
1615: (ARC-II, to \214nd the global minimum.)-1.012 E F1 2.75(4.2.3. Dynamic)
1616: 72 664.178 R(hill climbing \(DHC\))2.75 E F0 1.377
1617: (Michael de la Maza posted notices to public electronic b)99.5 690.078 R
1618: 1.377(ulletin boards, e.g., as summarized in a)-.22 F 3.397
1619: (public mailing list GA-List@AIC.NRL.N)72 712.078 R -1.485(AV)-.385 G
1620: -1.419(Y.)1.485 G 3.396(MIL, that his ne)1.419 F 6.146(wa)-.275 G 3.396
1621: (lgorithm, dynamic hill climbing)-6.146 F 1.407
1622: (\(DHC\), clearly outperformed genetic algorithms and ASA.)72 734.078 R
1623: 1.408(His code is a)6.908 F -.275(va)-.22 G 1.408
1624: (ilable by sending e-mail to).275 F EP
1625: %%Page: 20 20
1626: %%BeginPageSetup
1627: BP
1628: %%EndPageSetup
1629: /F0 11/Times-Roman@0 SF(ASA Lessons Learned)72 58 Q 163.138(-20- Lester)
1630: 125.859 F(Ingber)2.75 E 2.877(dhc@ai.mit.edu. Since)72 94 R .126
1631: (DHC is a v)2.877 F .126(ariant of a \231greedy\232 algorithm, it seeme\
1632: d appropriate to permit ASA to)-.275 F 1.203
1633: (also enter its quenching \(SQ\) domain.)72 116 R 1.203(The follo)6.703
1634: F 1.203(wing e)-.275 F 1.204(xcerpt is the reply posting in the abo)
1635: -.165 F 1.534 -.165(ve b)-.165 H(ulletin)-.055 E(board.)72 138 Q .942
1636: (\231SQ techniques lik)99.5 163.9 R 3.692(eG)-.11 G 3.691(Ao)-3.692 G
1637: -.165(bv)-3.691 G .941
1638: (iously are important and are crucial to solving man).165 F 3.691(ys)
1639: -.165 G .941(ystems in time)-3.691 F .392
1640: (periods much shorter than might be obtained by SA.)72 185.9 R .392
1641: (In ASA, if annealing is forsak)5.892 F .393(en, and Q)-.11 F(UENCHing)
1642: -.11 E 2.001(turned on, v)72 207.9 R 2.001(oiding the proof of sampling\
1643: , remarkable increases of speed can be obtained, apparently)-.22 F .577
1644: (sometimes e)72 229.9 R -.165(ve)-.275 G 3.327(ng).165 G .577
1645: (reater than other `greedy' algorithms.)-3.327 F -.165(Fo)6.077 G 3.327
1646: (re).165 G .577(xample, in \(Ingber and Rosen, 1992\) along)-3.492 F
1647: .638(with 5 GA test problems from the UCSD GA archi)72 251.9 R -.165(ve)
1648: -.275 G 3.387(,a).165 G .637
1649: (nother harder problem \(the ASA_TEST problem)-3.387 F 1.287
1650: (that comes with the ASA code\) w)72 273.9 R 1.287(as used.)-.11 F 1.287
1651: (In \(Ingber)6.787 F 4.037(,1)-.44 G 1.287(993b\) Q)-4.037 F 1.288
1652: (UENCHing w)-.11 F 1.288(as applied to this harder)-.11 F 2.896
1653: (problem. The)72 295.9 R .146(resulting SQ code w)2.896 F .146(as sho)
1654: -.11 F .145(wn to speed up the search by as much as as f)-.275 F .145
1655: (actor of 86 \(without)-.11 F -2.365 -.275(ev e)72 317.9 T 3.703(na).275
1656: G .953(ttempting to see if this could be increased further with more e)
1657: -3.703 F .953(xtreme quenching\).)-.165 F .954(This is greater)6.454 F
1658: .155(than the f)72 339.9 R .155(actor of 30 that w)-.11 F .154
1659: (as reported by Michael de la Maza for Dynamic Hill Climbing \(DHC\).)
1660: -.11 F .154(This is)5.654 F 4.56(as)72 361.9 S 1.81
1661: (imple change of one number in the code, turning it into a v)-4.56 F
1662: 1.811(ariant of SQ, and is not equi)-.275 F -.275(va)-.275 G 1.811
1663: (lent to).275 F 1.115(`tuning' an)72 383.9 R 3.865(yo)-.165 G 3.865(ft)
1664: -3.865 G 1.115(he other man)-3.865 F 3.864(yA)-.165 G 1.114
1665: (SA options, e.g., lik)-3.864 F 3.864(eS)-.11 G 1.114
1666: (ELF_OPTIMIZE, USER_COST_SCHEDULE,)-3.864 F 2.792(etc. Note)72 405.9 R
1667: .042(that SQ will not suf)2.792 F .042(\214ce for all systems; se)-.275
1668: F -.165(ve)-.275 G .043(ral users of ASA reported that Q).165 F .043
1669: (UENCHing did not)-.11 F(\214nd the global optimal point that w)72 427.9
1670: Q(as otherwise be found using the `correct' ASA algorithm.)-.11 E<9a>
1671: -.77 E/F1 11/Times-Bold@0 SF 2.75(4.2.4. Tsallis)72 471.9 R(statistics)
1672: 2.75 E F0 8.497(Ar)99.5 497.8 S 5.746(ecent paper claimed that a statis\
1673: tics whose parameterization permits an asymptotic)-8.497 F 1.471
1674: (approximation to the e)72 519.8 R 1.471(xponential function used for t\
1675: he Boltzmann of the standard SA acceptance test,)-.165 F .98(Tsallis st\
1676: atistics, is superior to the Boltzmann test \(Penna, 1994\), and an e)72
1677: 541.8 R .98(xample w)-.165 F .98(as gi)-.11 F -.165(ve)-.275 G 3.73(nc)
1678: .165 G(omparing)-3.73 E .885(standard SA to this ne)72 563.8 R 3.635(wa)
1679: -.275 G .885(lgorithm in the tra)-3.635 F -.165(ve)-.22 G .885
1680: (ling salesman problem \(TSP\).).165 F .885(There are tw)6.385 F 3.636
1681: (oi)-.11 G .886(ssues here:)-3.636 F(\(a\) the v)72 585.8 Q(alue of the\
1682: Tsallis test vs the Boltzmann test, and \(b\) the use of TSP for the c\
1683: on\214rmation of \(a\).)-.275 E 1.067(It seems v)99.5 611.7 R 1.067(ery\
1684: reasonable that the Tsallis test should be better than the Boltzmann t\
1685: est for the SA)-.165 F 1.64(acceptance test.)72 633.7 R -.165(Fo)7.14 G
1686: 4.39(re).165 G 1.641
1687: (xample, if the Boltzmann statistics did well on a gi)-4.555 F -.165(ve)
1688: -.275 G 4.391(nc).165 G 1.641(ost function)-4.391 F/F2 11/Times-Italic@0
1689: SF(C)4.215 E F0 4.391(,t).792 G 1.641(hen it)-4.391 F .519
1690: (might be the case that for the cost function)72 655.7 R F2(C)3.093 E/F3
1691: 11/Symbol SF 3.08<a23d>.792 G F0 -.165(ex)C(p\().165 E F2(C)-.176 E F0
1692: 3.269(\)am).792 G .519(ore moderate test, such as obtained for some)
1693: -3.269 F 2.283(parameterizations of the Tsallis statistics, w)72 677.7 R
1694: 2.283(ould be more appropriate to a)-.11 F -.22(vo)-.22 G 2.283
1695: (id getting stuck in local).22 F .493(minima of)72 699.7 R F2(C)3.067 E
1696: F3<a2>.792 E F0 5.993(.I)C 3.243(nf)-5.993 G .493
1697: (act, from its \214rst inception VFSR and ASA ha)-3.353 F .822 -.165
1698: (ve i)-.22 H .492(ncluded parameters to ef).165 F .492(fect similar)
1699: -.275 F(alternati)72 721.7 Q -.165(ve)-.275 G .696(s, and the latest v)
1700: .165 F .696(ersions of ASA no)-.165 F 3.446(wh)-.275 G -2.475 -.22(av e)
1701: -3.446 H .696(the Tsallis statistics as another alternati)3.666 F 1.026
1702: -.165(ve t)-.275 H .697(hat can).165 F EP
1703: %%Page: 21 21
1704: %%BeginPageSetup
1705: BP
1706: %%EndPageSetup
1707: /F0 11/Times-Roman@0 SF(ASA Lessons Learned)72 58 Q 163.138(-21- Lester)
1708: 125.859 F(Ingber)2.75 E .203(be commented out.)72 94 R 2.952(Ih)5.703 G
1709: -2.475 -.22(av e)-2.952 H .202(not yet e)3.172 F .202(xperienced an)
1710: -.165 F 2.952(ya)-.165 G(dv)-2.952 E .202(antages of this o)-.275 F
1711: -.165(ve)-.165 G 2.952(rt).165 G .202(he Boltzmann test when other)
1712: -2.952 F .03(ASA alternati)72 116 R -.165(ve)-.275 G 2.78(sa).165 G .031
1713: (re permitted to be used, b)-2.78 F .031(ut it seems lik)-.22 F .031
1714: (ely that there do e)-.11 F .031(xist some problems that might)-.165 F
1715: (bene\214t by its use.)72 138 Q 1.746(The use of TSP as a test for comp\
1716: arisons among SA techniques seems quite inappropriate.)99.5 163.9 R -.88
1717: (To)7.245 G .188(quote another source \(W)72 185.9 R .188
1718: (olpert and Macready)-.88 F 2.938(,1)-.715 G .188(995\): \231)-2.938 F
1719: .189(As an e)-.88 F .189(xample of this, it is well kno)-.165 F .189
1720: (wn that generic)-.275 F 1.44(methods \(lik)72 207.9 R 4.19(es)-.11 G
1721: 1.44(imulated annealing and genetic algorithms\) are unable to compete \
1722: with carefully hand-)-4.19 F 1.455
1723: (crafted solutions for speci\214c search problems.)72 229.9 R 1.456
1724: (The T)6.956 F(ra)-.385 E -.165(ve)-.22 G 1.456
1725: (ling Salesman \(TSP\) Problem is an e).165 F(xcellent)-.165 E -.165(ex)
1726: 72 251.9 S 2.032(ample of such a situation; the best search algorithms \
1727: for the TSP problem are hand-tailored for it).165 F(\(Reinelt, 1994\) .)
1728: 72 273.9 Q<9a>-.77 E/F1 11/Times-Bold@0 SF 2.75(4.2.5. Shubert)72 317.9
1729: R(pr)2.75 E(oblem)-.198 E F0 1.12(In \(Cvijo)99.5 343.8 R(vic)-.165 E
1730: <b4>-4.273 -.11 M 1.12(and Klino)4.48 .11 N 1.12(wski, 1995\) some stro\
1731: ng claims about the superiority of taboo search o)-.275 F -.165(ve)-.165
1732: G(r).165 E(genetic algorithms and simulated annealing were made.)72
1733: 365.8 Q 3.221(It)99.5 391.7 S .471(ook the only function gi)-3.221 F
1734: -.165(ve)-.275 G 3.22(ni).165 G 3.22(nt)-3.22 G .47
1735: (hat paper that also g)-3.22 F -2.475 -.22(av e)-.055 H .47
1736: (all parameters in this article, the Shubert)3.44 F 2.23
1737: (function, and ran a straight ASA test with absolutely no tuning.)72
1738: 413.7 R 2.23(\(I do consider tuning essential for)7.73 F 1.59
1739: (optimum use of an)72 435.7 R 4.339(yg)-.165 G 1.589
1740: (lobal optimization algorithm.\))-4.339 F -.88(Ta)7.089 G 1.589
1741: (ble 3 of that paper states that the results were).88 F 2.561(obtained \
1742: by counting the number of generated cost functions needed to \214rst at\
1743: tain a kno)72 457.7 R 2.562(wn global)-.275 F .425
1744: (minimum of -186.7309.)72 479.7 R .425
1745: (\(There are 18 global minima out of 760 local minima\).)5.925 F .424
1746: (That paper gi)5.924 F -.165(ve)-.275 G 3.174(sr).165 G(esults)-3.174 E
1747: 1.06(for taboo search \(TS\), and tw)72 501.7 R 3.81(ov)-.11 G 1.061
1748: (ariants of simulated annealing, SA1 and SA2, a)-4.085 F -.165(ve)-.22 G
1749: 1.061(raged o).165 F -.165(ve)-.165 G 3.811(r1).165 G 1.061(00 runs.)
1750: -3.811 F .703(SA1 required 241,215 generated cost functions; SA2 requir\
1751: ed 780 generated cost functions; TS required)72 523.7 R 2.207
1752: (727 generated cost functions.)72 545.7 R 2.207(The same test w)7.707 F
1753: 2.207(as performed with ASA, and it w)-.11 F 2.208(as found that ASA)
1754: -.11 F .807(required 577 generated cost functions.)72 567.7 R .806
1755: (More details of the coding of this problem and the results of the)6.307
1756: F(ASA calculation are gi)72 589.7 Q -.165(ve)-.275 G 2.75(ni).165 G 2.75
1757: (nt)-2.75 G(he NO)-2.75 E(TES \214le of the ASA code.)-.44 E .452
1758: (The result of course will v)99.5 615.6 R .453(ary some)-.275 F .453
1759: (what with the initial seed for the random number generator and)-.275 F
1760: .176(with the initial guess for the 2 parameters.)72 637.6 R .176(As ar)
1761: 5.676 F .176(gued for man)-.198 F 2.926(yo)-.165 G .176
1762: (ther cases, the f)-2.926 F .175(at tail of ASA, creating a)-.11 F .596
1763: (rather lar)72 659.6 R .597
1764: (ge generated to acceptance ratio, usually insures that the v)-.198 F
1765: .597(ariance of the results is not too lar)-.275 F(ge.)-.198 E 1.222(\(\
1766: Note that here there are 18 global minima which may cause more of a spr\
1767: ead.\))72 681.6 R 1.221(The 100 iterations of)6.722 F .261
1768: (ASA used the def)72 703.6 R .262
1769: (ault to let ASA randomly select the initial v)-.11 F .262
1770: (alues for the annealing loop \(as w)-.275 F .262(as implied)-.11 F
1771: 2.568(for SA1, SA2, and TS\).)72 725.6 R 2.568(The results from ASA \(s\
1772: imilar results not reported in that paper\) were a)8.068 F EP
1773: %%Page: 22 22
1774: %%BeginPageSetup
1775: BP
1776: %%EndPageSetup
1777: /F0 11/Times-Roman@0 SF(ASA Lessons Learned)72 58 Q 163.138(-22- Lester)
1778: 125.859 F(Ingber)2.75 E 1.524(minimum number of generated point to be 2\
1779: 31, a maximum number of generated point to be 1352, a)72 94 R
1780: (mean of 577.1, and a standard de)72 116 Q(viation of 145.6.)-.275 E/F1
1781: 11/Times-Bold@0 SF 2.75(4.2.6. Colville)72 160 R(pr)2.75 E(oblems)-.198
1782: E F0 .326(In early 1993 I w)99.5 185.9 R .326(as contacted by Zbignie)
1783: -.11 F 3.076(wM)-.275 G(ichale)-3.076 E .325(wicz, author of a ne)-.275
1784: F 3.075(wc)-.275 G .325(ode, GENECOP)-3.075 F -3.075 3.075(,a G)-1.221 H
1785: (A-)-3.075 E 1.658(based optimization package a)72 207.9 R -.275(va)-.22
1786: G 1.658(ilable via anon).275 F 1.659
1787: (ymous ftp from unccsun.uncc.edu in directory coe/e)-.165 F -.22(vo)
1788: -.275 G(l.).22 E .757(He stated that the ASA code f)72 229.9 R .757
1789: (ailed to get an)-.11 F 3.507(yr)-.165 G .757(easonable \214t to tw)
1790: -3.507 F 3.507(oC)-.11 G .757(olville problems, a set of problems)-3.507
1791: F(de)72 251.9 Q -.165(ve)-.275 G 1.975(loped by Colville, Floudas and P)
1792: .165 F 4.726(ardalos. He)-.165 F 1.976(demonstrated this by of)4.726 F
1793: 1.976(fering his o)-.275 F 1.976(wn results with)-.275 F .472
1794: (GENECOP in which he obtained better results that that obtained pre)72
1795: 273.9 R .472(viously in the literature on one these)-.275 F(problems.)72
1796: 295.9 Q 2.017(On one problem, after I learned that GENECOP \214rst made\
1797: algebraic substitutions to eliminate)99.5 321.8 R 1.608
1798: (equality constraints, this also w)72 343.8 R 1.607
1799: (as performed this before using ASA/VFSR.)-.11 F 1.607
1800: (GENECOP had produced)7.107 F 1.127(better results than pre)72 365.8 R
1801: 1.128(viously obtained in the literature for this e)-.275 F 1.128
1802: (xample, and ASA/VFSR obtained e)-.165 F -.165(ve)-.275 G(n).165 E 3.434
1803: (better results than GENECOP)72 387.8 R 8.934(.T)-1.221 G 3.434(he pre)
1804: -8.934 F 3.434(viously published results g)-.275 F -2.475 -.22(av e)
1805: -.055 H 6.184<618c>6.404 G 3.434(nal cost of -47.707579;)-6.184 F 2.213
1806: (GENECOP g)72 409.8 R -2.475 -.22(av e)-.055 H 2.213
1807: (cost of -47.760765; ASA/VFSR g)5.183 F -2.475 -.22(av e)-.055 H 2.214
1808: (cost of -47.76109.)5.184 F -.165(Fo)7.714 G 4.964(rs).165 G 2.214
1809: (uch systems, I ha)-4.964 F -.165(ve)-.22 G
1810: (suggested that GENECOP be used as a front end to ASA.)72 431.8 Q 1.049
1811: (On the other problem, using the def)99.5 457.7 R 1.048
1812: (ault parameters of ASA, the code did indeed ho)-.11 F -.165(ve)-.165 G
1813: 3.798(ri).165 G 3.798(nal)-3.798 G(ocal)-3.798 E 2.356(minimum, lik)72
1814: 479.7 R 2.356(ely for so long that it might tak)-.11 F 5.106(ea)-.11 G
1815: 5.106(ne)-5.106 G 2.357(normous amount of human time and unrealistic)
1816: -5.106 F 2.165(computer precision \(to k)72 501.7 R 2.165(eep sampling \
1817: at incredibly small temperatures\) to \214nd the global minimum.)-.11 F
1818: (Ho)72 523.7 Q(we)-.275 E -.165(ve)-.275 G 2.171 -.44(r, j).165 H 1.291
1819: (ust a fe).44 F 4.041(wm)-.275 G 1.292(inutes of modestly tuning a fe)
1820: -4.041 F 4.042(wA)-.275 G 1.292
1821: (SA parameters produced the global minimum)-4.042 F -.165(ve)72 545.7 S
1822: (ry ef).165 E(\214ciently)-.275 E(.)-.715 E .408
1823: (More details of the coding of these tw)99.5 571.6 R 3.157(op)-.11 G
1824: .407(roblems and the results of the ASA calculations are gi)-3.157 F
1825: -.165(ve)-.275 G(n).165 E(in the NO)72 593.6 Q
1826: (TES \214le of the ASA code.)-.44 E F1 2.75(5. CONCLUSION)72 637.6 R F0
1827: .09(If ask)99.5 663.5 R .09(ed to state one major common feature of non\
1828: linear systems, in the conte)-.11 F .091(xt of optimization, the)-.165 F
1829: .571(feature most lik)72 685.5 R .571(ely should be gi)-.11 F -.165(ve)
1830: -.275 G 3.321(ni).165 G 3.321(st)-3.321 G .571
1831: (hat nonlinear systems typically are non-typical.)-3.321 F .571
1832: (It is unlik)6.071 F .57(ely that)-.11 F(an)72 707.5 Q 3.072<7999>-.165
1833: G .322(canned black-box\232 code can be de)-3.072 F -.165(ve)-.275 G
1834: .322(loped, requiring no or fe).165 F 3.072(wm)-.275 G .323
1835: (inor adjustments, that will usefully)-3.072 F 2.096(guarantee ef)72
1836: 729.5 R 2.096(\214cient global optimization for se)-.275 F -.165(ve)
1837: -.275 G 2.095(rely nonlinear systems, e.g., similar to what might be)
1838: .165 F EP
1839: %%Page: 23 23
1840: %%BeginPageSetup
1841: BP
1842: %%EndPageSetup
1843: /F0 11/Times-Roman@0 SF(ASA Lessons Learned)72 58 Q 163.138(-23- Lester)
1844: 125.859 F(Ingber)2.75 E -.165(ex)72 94 S(pected for man).165 E 2.75(yq)
1845: -.165 G(uasi-linear systems.)-2.75 E 1.818(In the absence of kno)99.5
1846: 119.9 R 1.818(wledge about a particular system, gi)-.275 F -.165(ve)
1847: -.275 G 4.569(nt).165 G 1.819(hat only SA can of)-4.569 F 1.819
1848: (fer at least a)-.275 F .802(\231statistical\232 proof of global optimi\
1849: zation, then the \214rst algorithm of choice clearly is SA.)72 141.9 R
1850: (Modi\214cations)6.301 E .362
1851: (of SA, e.g., SQ quenching algorithms, may be competiti)72 163.9 R .692
1852: -.165(ve w)-.275 H .362(ith other techniques, e.g., simple).165 F 3.112
1853: (xo)-.165 G 3.112(rg)-3.112 G(enetic)-3.112 E .304(algorithms, b)72
1854: 185.9 R .304(ut among these the best choice is not so clear)-.22 F 5.804
1855: (.S)-.605 G 3.054(Qd)-5.804 G .304(oes of)-3.054 F .304(fer a relati)
1856: -.275 F -.165(ve)-.275 G .304(ly simple approach to).165 F
1857: (quickly writing code for optimization, b)72 207.9 Q
1858: (ut ultimately the end results must justify this means.)-.22 E .687
1859: (This ar)99.5 233.8 R .687
1860: (gument for the use of SA has an opposite side.)-.198 F .687
1861: (If some information about a system can be)6.187 F .891(incorporated in\
1862: to some other global optimization technique, and it can be determined t\
1863: hat the technique)72 255.8 R .667(can deli)72 277.8 R -.165(ve)-.275 G
1864: 3.417(rt).165 G .667
1865: (he global optimum point, often that technique will be more ef)-3.417 F
1866: .668(\214cient than SA.)-.275 F .668(E.g., a quasi-)6.168 F(Ne)72 299.8
1867: Q(wton algorithm will be more ef)-.275 E
1868: (\214cient than SA for parabolic systems.)-.275 E -.165(Fo)99.5 325.7 S
1869: 4.4(rm).165 G(an)-4.4 E 4.4(yr)-.165 G 1.65(esearchers, the \214rst cho\
1870: ice of algorithm to use for a nonlinear or stochastic problem)-4.4 F
1871: (lik)72 347.7 Q 1.374(ely will be one with which the)-.11 F 4.124(ya)
1872: -.165 G 1.374(lready are f)-4.124 F(amiliar)-.11 E 6.874(.I)-.605 G
1873: 4.124(ft)-6.874 G 1.374(hat f)-4.124 F 1.374
1874: (ails, then SA is an option to try ne)-.11 F(xt.)-.165 E 1.769
1875: (More research needs to be done to see if a more objecti)72 369.7 R
1876: 2.099 -.165(ve c)-.275 H 1.768
1877: (lassi\214cation of nonlinear systems can be).165 F(de)72 391.7 Q -.165
1878: (ve)-.275 G 1.136(loped to help guide a gi).165 F -.165(ve)-.275 G 3.886
1879: (nr).165 G 1.136(esearcher to a gi)-3.886 F -.165(ve)-.275 G 3.887(na)
1880: .165 G 1.137(lgorithm for a gi)-3.887 F -.165(ve)-.275 G 3.887(np).165 G
1881: 3.887(roblem. As)-3.887 F 1.137(the e)3.887 F(xamples)-.165 E .158
1882: (included in the documentation of the ASA code illustrate, there ha)72
1883: 413.7 R .488 -.165(ve b)-.22 H .158
1884: (een \231surprises\232 whereby some v).165 F(ery)-.165 E(dif)72 435.7 Q
1885: .283(\214cult problems ha)-.275 F .613 -.165(ve b)-.22 H .283
1886: (een quickly solv).165 F .284(ed by ASA, while others ha)-.165 F .614
1887: -.165(ve r)-.22 H .284(equired quite a bit of \231tuning\232 to).165 F
1888: (establish a good set of starting OPTIONS.)72 457.7 Q 3.239(Especially \
1889: among \214rst-time users of SA, often there is much misunderstanding an\
1890: d lack of)99.5 483.6 R 1.22(appreciation of just what an SA code can im\
1891: mediately do for a particular problem.)72 505.6 R 1.22
1892: (Some education is)6.72 F .598(necessary to mak)72 527.6 R 3.347(eu)-.11
1893: G .597(sers a)-3.347 F -.11(wa)-.165 G .597(re of the potential problem\
1894: s that may arise and what remedies the particular).11 F
1895: (algorithm can of)72 549.6 Q(fer to o)-.275 E -.165(ve)-.165 G
1896: (rcome these obstacles.).165 E EP
1897: %%Page: 24 24
1898: %%BeginPageSetup
1899: BP
1900: %%EndPageSetup
1901: /F0 11/Times-Roman@0 SF(ASA Lessons Learned)72 58 Q 163.138(-24- Lester)
1902: 125.859 F(Ingber)2.75 E/F1 11/Times-Bold@0 SF(REFERENCES)72 94 Q F0(B)72
1903: 123.8 Q/F2 9/Times-Roman@0 SF(INDER)A F0 4.807(,K)C(.)-4.807 E F2(AND)
1904: 4.807 E F0(S)4.807 E F2 -1.071 -.837(TA U)D(FFER).837 E F0 4.808(,D)C
1905: 2.058
1906: (., \(1985\) A simple introduction to Monte Carlo simulations and some)
1907: -4.808 F 1.502(specialized topics, In: Applications of the Monte Carlo \
1908: Method in Statistical Ph)99.5 145.8 R 1.501(ysics, K. Binder)-.055 F
1909: (\(Ed.\), Berlin, Springer)99.5 167.8 Q(-V)-.22 E(erlag.)-1.221 E(B)72
1910: 193.7 Q F2 -1.845 -.36(RO W)D(N).36 E F0 2.815(,R)C .065(., R)-2.815 F
1911: F2(ULK)A -.45(OV)-.27 G F0 2.815(,N).45 G(.F)-2.815 E(.,)-.88 E F2(AND)
1912: 2.815 E F0(T)2.815 E F2(UFILLAR)A(O)-.36 E F0 2.815(,N)C .066
1913: (.B., \(1994\) Synchronization of chaotic systems: the ef)-2.815 F
1914: (fects)-.275 E(of additi)99.5 215.7 Q .33 -.165(ve n)-.275 H
1915: (oise and drift in the dynamics and dri).165 E(ving, Ph)-.275 E(ys. Re)
1916: -.055 E 1.43 -.715(v. E \()-.275 H(to be published\).).715 E(C)72 241.6
1917: Q F2(ERNY)A F0 6.149(,V)C 3.399
1918: (., \(1982\) A thermodynamical approach to the tra)-7.568 F -.165(ve)
1919: -.22 G 3.398(lling salesman problem: An ef).165 F(\214cient)-.275 E
1920: (simulation algorithm, Report, Bratisla)99.5 263.6 Q -.275(va)-.22 G
1921: 2.75(,C).275 G(zechoslo)-2.75 E -.275(va)-.165 G(kia, Comenius Uni).275
1922: E -.165(ve)-.275 G(rsity).165 E(.)-.715 E(C)72 289.5 Q F2(OHEN)A F0
1923: 3.473(,B)C .723(., \(1994\) T)-3.473 F .723
1924: (raining synaptic delays in a recurrent neural netw)-.385 F .723
1925: (ork, M.S. Thesis, T)-.11 F(el-A)-.77 E(vi)-.814 E 2.154 -.715(v, I)
1926: -.275 H(srael,).715 E -.77(Te)99.5 311.5 S(l-A).77 E(vi)-.814 E 2.75(vU)
1927: -.275 G(ni)-2.75 E -.165(ve)-.275 G(rsity).165 E(.)-.715 E(C)72 337.4 Q
1928: F2(ORAN)A(A)-.315 E F0 3.711(,A)C .961(., M)-3.711 F F2(ARCHESI)A F0
1929: 3.711(,M)C .96(., M)-3.711 F F2(AR)A(TINI)-.54 E F0 3.71(,C)C(.,)-3.71 E
1930: F2(AND)3.71 E F0(R)3.71 E F2(IDELLA)A F0 3.71(,S)C .96
1931: (., \(1987\) Minimizing multimodal functions)-3.71 F .263
1932: (of continuous v)99.5 359.4 R .263
1933: (ariables with the "simulated annealing" algorithm, A)-.275 F .264(CM T)
1934: -.44 F .264(rans. Mathl. Softw)-.385 F(are)-.11 E F1(13)3.014 E F0(,)A
1935: (3, 262-279.)99.5 381.4 Q(C)72 407.3 Q F2(VIJO)A(VIC)-.45 E<b4>-4.5
1936: -2.034 M F0 3.956(,D)1.503 2.034 O(.)-3.956 E F2(AND)3.956 E F0(K)3.956
1937: E F2(LINO)A(WSKI)-.315 E F0 3.956(,J)C 1.206(., \(1995\) T)-3.956 F
1938: 1.205(aboo search: An approach to the multiple minima problem,)-.88 F
1939: (Science)99.5 429.3 Q F1(267)2.75 E F0 2.75(,5)C(198, 664-666.)-2.75 E
1940: (D)72 455.2 Q F2(E)A F0(J)5.5 E F2(ONG)A F0 4.711(,K)C 1.961
1941: (.A., \(1992\) Genetic algorithms are NO)-4.711 F 4.711(Tf)-.44 G 1.961
1942: (unction optimizers, In: F)-4.711 F 1.961(oundations of Genetic)-.165 F
1943: (Algorithms: Proceedings 24-29 July 1992, D. Whitle)99.5 477.2 Q 2.75
1944: (y\()-.165 G(Ed.\), V)-2.75 E(ail, CO, Mor)-1.221 E -.055(ga)-.198 G
1945: 2.75(nK).055 G(aufman.)-2.75 E(F)72 503.1 Q F2(ORREST)A F0 3.637(,S)C
1946: .887(., \(1993\) Genetic algorithms: Principles of natural selection ap\
1947: plied to computation, Science)-3.637 F F1(261)99.5 525.1 Q F0 2.75(,5)C
1948: (123, 872-878.)-2.75 E(F)72 551 Q F2 -.36(RO)C(ST).36 E F0 3.823(,R)C
1949: 1.073(., \(1993\) Ensemble Based Simulated Annealing \(EBSA\), ftp.sdsc\
1950: .edu:/pub/sdsc/math/Ebsa, La)-3.823 F(Jolla, CA, Uni)99.5 573 Q -.165
1951: (ve)-.275 G(rsity of California San Die).165 E(go.)-.165 E(G)72 598.9 Q
1952: F2(EMAN)A F0 2.85(,S)C(.)-2.85 E F2(AND)2.85 E F0(G)2.85 E F2(EMAN)A F0
1953: 2.85(,D)C .1(., \(1984\) Stochastic relaxation, Gibbs distrib)-2.85 F
1954: .099(ution and the Bayesian restoration)-.22 F(in images, IEEE T)99.5
1955: 620.9 Q(rans. P)-.385 E(att. Anal. Mac. Int.)-.165 E F1(6)5.5 E F0 2.75
1956: (,6)C 2.75(,7)-2.75 G(21-741.)-2.75 E(I)72 646.8 Q F2(NDIVERI)A F0 2.973
1957: (,G)C .223(., N)-2.973 F F2 -.999(AT)C(ERI).999 E F0 2.973(,G)C .223
1958: (., R)-2.973 F F2(AFFO)A F0 2.974(,L)C(.,)-2.974 E F2(AND)2.974 E F0(C)
1959: 2.974 E F2 -1.215(AV)C(IGLIA)1.215 E F0 2.974(,D)C .224
1960: (., \(1993\) A neural netw)-2.974 F .224(ork architecture for defect)
1961: -.11 F(detection through magnetic inspection, Report, Geno)99.5 668.8 Q
1962: -.275(va)-.165 G 2.75(,I).275 G(taly)-2.75 E 2.75(,U)-.715 G(ni)-2.75 E
1963: -.165(ve)-.275 G(rsity of Geno).165 E -.275(va)-.165 G(.).275 E(I)72
1964: 694.7 Q F2(NGBER)A F0 2.75(,L)C(., \(1989\) V)-2.75 E(ery f)-1.221 E
1965: (ast simulated re-annealing, Mathl. Comput. Modelling)-.11 E F1(12)2.75
1966: E F0 2.75(,8)C 2.75(,9)-2.75 G(67-973.)-2.75 E EP
1967: %%Page: 25 25
1968: %%BeginPageSetup
1969: BP
1970: %%EndPageSetup
1971: /F0 11/Times-Roman@0 SF(ASA Lessons Learned)72 58 Q 163.138(-25- Lester)
1972: 125.859 F(Ingber)2.75 E(I)72 94 Q/F1 9/Times-Roman@0 SF(NGBER)A F0 3.288
1973: (,L)C .537(., \(1990\) Statistical mechanical aids to calculating term \
1974: structure models, Ph)-3.288 F .537(ys. Re)-.055 F 1.967 -.715(v. A)-.275
1975: H/F2 11/Times-Bold@0 SF(42)4.002 E F0 3.287(,1)C(2,)-3.287 E(7057-7064.)
1976: 99.5 116 Q(I)72 141.9 Q F1(NGBER)A F0 5.474(,L)C 2.725(., \(1991\) Stat\
1977: istical mechanics of neocortical interactions: A scaling paradigm appli\
1978: ed to)-5.474 F(electroencephalograph)99.5 163.9 Q 1.43 -.715(y, P)-.055
1979: H -.055(hy).715 G(s. Re).055 E 1.43 -.715(v. A)-.275 H F2(44)3.465 E F0
1980: 2.75(,6)C 2.75(,4)-2.75 G(017-4060.)-2.75 E(I)72 189.8 Q F1(NGBER)A F0
1981: 4.469(,L)C 1.719(., \(1992\) Generic mesoscopic neural netw)-4.469 F
1982: 1.718(orks based on statistical mechanics of neocortical)-.11 F
1983: (interactions, Ph)99.5 211.8 Q(ys. Re)-.055 E 1.43 -.715(v. A)-.275 H F2
1984: (45)3.465 E F0 2.75(,4)C 2.75(,R)-2.75 G(2183-R2186.)-2.75 E(I)72 237.7
1985: Q F1(NGBER)A F0 4.752(,L)C 2.003(., \(1993a\) Adapti)-4.752 F 2.333
1986: -.165(ve S)-.275 H 2.003
1987: (imulated Annealing \(ASA\), [ftp.alumni.caltech.edu: /pub/ingber/ASA-)
1988: .165 F(shar)99.5 259.7 Q 2.75(,A)-.44 G(SA-shar)-2.75 E(.Z, ASA.tar)
1989: -.605 E(.Z, ASA.tar)-.605 E(.gz, ASA.zip], McLean, V)-.605 E
1990: (A, Lester Ingber Research.)-1.485 E(I)72 285.6 Q F1(NGBER)A F0 4.96(,L)
1991: C 2.21(., \(1993b\) Simulated annealing: Practice v)-4.96 F 2.21
1992: (ersus theory)-.165 F 4.96(,M)-.715 G 2.21(athl. Comput. Modelling)-4.96
1993: F F2(18)4.959 E F0 4.959(,1)C(1,)-4.959 E(29-57.)99.5 307.6 Q(I)72 333.5
1994: Q F1(NGBER)A F0 5.578(,L)C 2.828(., F)-5.578 F F1(UJIO)A F0 5.578(,H)C
1995: (.,)-5.578 E F1(AND)5.578 E F0(W)5.578 E F1(EHNER)A F0 5.578(,M)C(.F)
1996: -5.578 E 2.828(., \(1991\) Mathematical comparison of combat computer)
1997: -.88 F(models to e)99.5 355.5 Q -.165(xe)-.165 G
1998: (rcise data, Mathl. Comput. Modelling).165 E F2(15)2.75 E F0 2.75(,1)C
1999: 2.75(,6)-2.75 G(5-90.)-2.75 E(I)72 381.4 Q F1(NGBER)A F0 6.844(,L)C(.)
2000: -6.844 E F1(AND)6.844 E F0(R)6.844 E F1(OSEN)A F0 6.844(,B)C 4.094
2001: (., \(1992\) Genetic algorithms and v)-6.844 F 4.094(ery f)-.165 F 4.093
2002: (ast simulated reannealing: A)-.11 F
2003: (comparison, Mathl. Comput. Modelling)99.5 403.4 Q F2(16)2.75 E F0 2.75
2004: (,1)C(1, 87-100.)-2.75 E(I)72 429.3 Q F1(NGBER)A F0 4.493(,L)C(.)-4.493
2005: E F1(AND)4.493 E F0(S)4.493 E F1 -.09(WO)C(RDER).09 E F0 4.493(,D)C
2006: 1.744(.D., \(1991\) Statistical mechanics of combat with human f)-4.493
2007: F 1.744(actors, Mathl.)-.11 F(Comput. Modelling)99.5 451.3 Q F2(15)2.75
2008: E F0 2.75(,1)C(1, 99-127.)-2.75 E(I)72 477.2 Q F1(NGBER)A F0 4.748(,L)C
2009: 1.998(., W)-4.748 F F1(EHNER)A F0 4.748(,M)C(.F)-4.748 E 1.998(., J)-.88
2010: F F1(ABBOUR)A F0 4.747(,G)C(.M.,)-4.747 E F1(AND)4.747 E F0(B)4.747 E F1
2011: (ARNHILL)A F0 4.747(,T)C 1.997(.M., \(1991\) Application of statistical)
2012: -5.561 F .571(mechanics methodology to term-structure bond-pricing mode\
2013: ls, Mathl. Comput. Modelling)99.5 499.2 R F2(15)3.322 E F0 3.322(,1)C
2014: (1,)-3.322 E(77-98.)99.5 521.2 Q(K)72 547.1 Q F1(IMURA)A F0 4.827(,K)C
2015: (.)-4.827 E F1(AND)4.827 E F0(T)4.827 E F1(AKI)A F0 4.827(,K)C 2.076
2016: (., \(1991\) T)-4.827 F 2.076
2017: (ime-homogeneous parallel annealing algorithm, Report TR-673,)-.385 F
2018: -.88(To)99.5 569.1 S -.165(ky).88 G(o, Japan, Institute for Ne).165 E
2019: 2.75(wG)-.275 G(eneration Computer T)-2.75 E(echnology)-.77 E(.)-.715 E
2020: (K)72 595 Q F1(IRKP)A -.999(AT)-.828 G(RICK).999 E F0 4.102(,S)C 1.352
2021: (., G)-4.102 F F1(ELA)A(TT)-.999 E F0 4.102(,C)C 1.352(.D., J)-4.102 F
2022: F1(R)A F0(.,)A F1(AND)4.102 E F0(V)4.102 E F1(ECCHI)A F0 4.103(,M)C(.P)
2023: -4.103 E 1.353(., \(1983\) Optimization by simulated annealing,)-1.221 F
2024: (Science)99.5 617 Q F2(220)2.75 E F0 2.75(,4)C(598, 671-680.)-2.75 E(M)
2025: 72 642.9 Q F1(A)A F0 2.75(,S)C
2026: (.-K., \(1985\) Statistical Mechanics, Philadelphia, W)-2.75 E
2027: (orld Scienti\214c.)-.88 E(M)72 668.8 Q F1 -.999(AT)C(HEWS).999 E F0
2028: 4.214(,J)C(.)-4.214 E F1(AND)4.214 E F0(W)4.214 E F1(ALKER)A F0 4.214
2029: (,R)C 1.463(.L., \(1970\) Mathematical Methods of Ph)-4.214 F 1.463
2030: (ysics, 2nd ed., Ne)-.055 F 4.213(wY)-.275 G 1.463(ork, NY)-5.423 F(,)
2031: -1.419 E(Benjamin.)99.5 690.8 Q(M)72 716.7 Q F1(ETR)A(OPOLIS)-.36 E F0
2032: 5.006(,N)C 2.256(., R)-5.006 F F1(OSENBLUTH)A F0 5.006(,A)C(.W)-5.006 E
2033: 2.256(., R)-1.012 F F1(OSENBLUTH)A F0 5.006(,M)C 2.257(.N., T)-5.006 F
2034: F1(ELLER)A F0 5.007(,A)C(.H.,)-5.007 E F1(AND)5.007 E F0(T)5.007 E F1
2035: (ELLER)A F0 5.007(,E)C 2.257(., \(1953\))-5.007 F
2036: (Equation of state calculations by f)99.5 738.7 Q
2037: (ast computing machines, J. Chem. Ph)-.11 E(ys.)-.055 E F2(21)5.5 E F0
2038: 2.75(,6)C 2.75(,1)-2.75 G(087-1092.)-2.75 E EP
2039: %%Page: 26 26
2040: %%BeginPageSetup
2041: BP
2042: %%EndPageSetup
2043: /F0 11/Times-Roman@0 SF(ASA Lessons Learned)72 58 Q 163.138(-26- Lester)
2044: 125.859 F(Ingber)2.75 E(P)72 94 Q/F1 9/Times-Roman@0 SF(ENN)A(A)-.315 E
2045: F0 2.75(,T)C(.J.P)-3.564 E(., \(1994\) T)-1.221 E(ra)-.385 E -.165(ve)
2046: -.22 G(ling salesman problem and Tsallis statistics, Ph).165 E(ys. Re)
2047: -.055 E 1.43 -.715(v. E)-.275 H/F2 11/Times-Bold@0 SF(50)3.465 E F0 2.75
2048: (,6)C 2.75(,R)-2.75 G(1-R3.)-2.75 E(P)72 119.9 Q F1(INCUS)A F0 3.636(,M)
2049: C .885(., \(1970\) A Monte Carlo method for the approximate solution of\
2050: certain types of constrained)-3.636 F(optimization problems, Oper)99.5
2051: 141.9 Q 2.75(.R)-.605 G(es.)-2.75 E F2(18)5.5 E F0 2.75(,1)C(225-1228.)
2052: -2.75 E(R)72 167.8 Q F1(EINEL)A(T)-.828 E F0 4.94(,G)C 2.191
2053: (., \(1994\) The T)-4.94 F(ra)-.385 E -.165(ve)-.22 G 2.191
2054: (ling Salesman, computational solutions for TSP applications, Berling,)
2055: .165 F(Springer)99.5 189.8 Q(.)-.605 E(R)72 215.7 Q F1(OSEN)A F0 4.346
2056: (,B)C 1.595(., \(1992\) Function optimization based on adv)-4.346 F
2057: 1.595(anced simulated annealing, IEEE W)-.275 F 1.595(orkshop on)-.88 F
2058: (Ph)99.5 237.7 Q(ysics and Computation - Ph)-.055 E(ysComp '92 289-293.)
2059: -.055 E(S)72 263.6 Q F1(CHRA)A(UDOLPH)-.495 E F0 3.003(,N)C(.N.)-3.003 E
2060: F1(AND)3.003 E F0(G)3.003 E F1(REFENSTETTE)A F0 3.004(,J)C .254
2061: (.J., \(1991\) A Users Guide to GA)-3.004 F .254
2062: (UCSD 1.2, Report, La Jolla,)-.605 F(CA, Uni)99.5 285.6 Q -.165(ve)-.275
2063: G(rsity of California at San Die).165 E(go.)-.165 E(S)72 311.5 Q F1(ZU)A
2064: F0 2.75(,H)C(.)-2.75 E F1(AND)2.75 E F0(H)2.75 E F1(AR)A(TLEY)-.54 E F0
2065: 2.75(,R)C(., \(1987\) F)-2.75 E(ast simulated annealing, Ph)-.165 E
2066: (ys. Lett. A)-.055 E F2(122)2.75 E F0 2.75(,3)C(-4, 157-162.)-2.75 E(T)
2067: 72 337.4 Q F1(ANG)A F0 4.604(,X)C 1.854(.Z., T)-4.604 F F1(RA)A(CY)-.36
2068: E F0 4.604(,E)C 1.854(.R., B)-4.604 F F1(OOZER)A F0 4.604(,A)C(.D.,)
2069: -4.604 E F1(DE)4.604 E F0(B)A F1(RA)A(UW)-.495 E F0 4.604(,A)C(.,)-4.604
2070: E F1(AND)4.603 E F0(B)4.603 E F1 -1.845 -.36(RO W)D(N).36 E F0 4.603(,R)
2071: C 1.853(., \(1995\) Symbol sequence)-4.603 F
2072: (statistics in noisy chaotic signal reconstruction, Ph)99.5 359.4 Q
2073: (ys. Re)-.055 E 1.43 -.715(v. E)-.275 H F2(51)3.465 E F0 2.75(,4)C 2.75
2074: (,\()-2.75 G(in press\).)-2.75 E F1 -1.215(VA)72 385.3 S(N)1.215 E F0(L)
2075: 5.5 E F1(AARHO)A(VEN)-.45 E F0 4.302(,P)C(.J.M.)-5.523 E F1(AND)4.302 E
2076: F0(A)4.302 E F1(AR)A(TS)-.54 E F0 4.302(,E)C 1.553
2077: (.H.L., \(1987\) Simulated Annealing: Theory and Applications,)-4.302 F
2078: (Dordrecht, The Netherlands, D. Reidel.)99.5 407.3 Q(W)72 433.2 Q F1
2079: (OFSEY)A F0 2.793(,M)C .043(., \(1993\) T)-2.793 F .043
2080: (echnology: Shortcut tests v)-.77 F .042
2081: (alidity of complicated formulas, The W)-.275 F .042(all Street Journal)
2082: -.88 F F2(222)99.5 455.2 Q F0 2.75(,6)C(0, B1.)-2.75 E(W)72 481.1 Q F1
2083: (OLPER)A(T)-.54 E F0 4.329(,D)C(.H.)-4.329 E F1(AND)4.329 E F0(M)4.329 E
2084: F1 -.36(AC)C(READ).36 E(Y)-.495 E F0 4.329(,W)C 1.579
2085: (.G., \(1995\) No free lunch theorems for search, Report, Santa Fe,)
2086: -5.341 F(NM, Santa Fe Institute.)99.5 503.1 Q(W)72 529 Q F1(U)A F0 4.189
2087: (,K)C(.)-4.189 E F1(AND)4.188 E F0(L)4.188 E F1(EVINE)A F0 4.188(,M)C
2088: 1.438(.D., \(1993\) 3-D object representation using parametric geons, T\
2089: R-CIM-93-13,)-4.188 F
2090: (Montreal, Canada, Center for Intelligent Machines, McGill Uni)99.5 551
2091: Q -.165(ve)-.275 G(rsity).165 E(.)-.715 E EP
2092: %%Trailer
2093: end
2094: %%EOF
2095: