1: \documentclass{llncs}
2:
3: %\usepackage{amsmath,amsfonts,amssymb}
4: %\usepackage{epsfig,psfig}
5: %\usepackage{graphicx}
6:
7: \usepackage[dvips]{graphicx}
8: \usepackage{isolatin1}
9: \bibliographystyle{plain}
10:
11: \begin{document}
12:
13: \mainmatter
14:
15:
16: \title{
17: Dominance Based Crossover Operator \\ %\medskip
18: for Evolutionary Multi-objective Algorithms\\
19: {\bf DRAFT VERSION}
20: }
21:
22: \vspace{0.5cm}
23:
24: \author
25: { Olga Rudenko, Marc Schoenauer}
26:
27: \vspace{0.3cm}
28: \institute{
29: TAO Team, INRIA Futurs \\
30: LRI, bat. 490, Université Paris-Sud\\
31: 91405 Orsay Cedex, France\\
32: Olga.Roudenko@lri.fr, Marc.Schoenauer@inria.fr
33: }
34:
35: \maketitle
36:
37: \pagestyle{empty}
38:
39: \begin{abstract}
40: \noindent
41:
42: In spite of the recent quick growth of the Evolutionary
43: Multi-objective Optimization (EMO) research field, there has been few
44: trials to adapt the general variation operators to the particular
45: context of the quest for the Pareto-optimal set. The only exceptions
46: are some mating restrictions that take in account the distance between
47: the potential mates -- but contradictory conclusions have been reported.
48: This paper introduces a particular mating restriction for Evolutionary
49: Multi-objective Algorithms, based on the Pareto dominance relation:
50: the partner of a non-dominated individual will be preferably chosen
51: among the individuals of the population that it dominates. Coupled
52: with the BLX crossover operator, two different ways of generating
53: offspring are proposed. This recombination scheme is validated within
54: the well-known NSGA-II framework on three bi-objective benchmark
55: problems and one real-world bi-objective constrained optimization
56: problem. An acceleration of the progress of the population toward the
57: Pareto set is observed on all problems.
58:
59: \end{abstract}
60:
61: \section*{Introduction}
62: %%%%%%%%%%%%%%%%%%%%%%%%%
63: The idea of {\em restricted mating} is not new in Evolutionary
64: Computation: Goldberg~\cite{Goldberg:89} already suggested to forbid,
65: or at least restrict, the crossover between too different individuals
66: (i.e. that are too far apart for some distance on the genotypic
67: space) -- which makes sense for single-objective problems as soon as
68: the population has started to accumulate on different fitness peaks,
69: as recombining individuals from different peaks would often
70: lead to lethal individuals.
71: This idea has been transposed in the framework of Evolutionary
72: Multi-objective Algorithms (EMAs) by
73: Hajela and Lin~\cite{HajelaLin:92}, and by Fonseca and Fleming~\cite{FF93}.
74: Nevertheless, Zitzler and Thiele~\cite{ZitzlerThiele:PPSN98} did not
75: observe any improvement when mating similar %(in the objective space)
76: individuals.
77: %
78: On the other hand, Horn et
79: al.~\cite{Horn:al:94} present an argument supporting mating of
80: dissimilar individuals: in the
81: multi-objective framework, because the population diversity is
82: enforced, the information provided by very different
83: solutions can be combined in such way that a new type of (good)
84: compromises can hopefully be obtained.
85: Nevertheless, Schaffer reported the absence of the
86: improvement when mating dissimilar individuals.
87: To sum up, no clear conclusion can be drawn from existing experiments
88: on the usefulness of restricted mating based on the (dis)similarity
89: between mates.
90:
91: On a real-world design problem, using a very specific
92: representation, Wildman et Parks~\cite{WildmanParks:EMO2} have
93: investigated different pairing strategies based on maximizing or
94: minimizing different similarity measures. In particular, the
95: similarity in the sense of the dominance rank has been considered, and
96: enforcing the mating of the individuals from the elite archive with
97: the individuals from the population, in an archive-based EMA, has been
98: observed to be beneficial.
99:
100: However, in all studies enumerated above, the efficiency of the
101: proposed mating restrictions has been measured by the
102: quality of the final non-dominated solutions, without addressing the
103: issue of computational time. In this paper, we propose a restricted
104: mating strategy whose main effect is to accelerate the progress of the
105: population of an EMA toward the Pareto set. The main idea is fairly
106: simple, and
107: consists in using the Pareto dominance relation when choosing a mate
108: for the best (non-dominated) individuals. Note that a more detailed
109: presentation (in French) can be found in \cite{these-Olga}.
110:
111: The paper is organized as follows. Next section briefly
112: introduces evolutionary multi-objective optimization, and describes
113: in more detail the NSGA-II algorithm, one of the best performing EMA
114: to-date, that will be used in all experiments. Two slightly
115: different implementations of the dominance-based crossover operator
116: are then proposed in Section \ref{xover}, based on BLX-$\alpha$
117: crossover, used throughout this study. Section \ref{experiments}
118: presents some
119: experimental results witnessing the acceleration of the progress
120: toward the Pareto set when using the proposed mating
121: restrictions. Finally, Section \ref{conclusion} gives some guidelines
122: for a more rigorous and
123: complete validation of the proposed strategy, as well
124: as for its possible refinements.
125:
126:
127: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
128: \section{Evolutionary Multi-objective Optimization}
129: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
130: \label{emo}
131:
132: Multi-objective optimization aims at simultaneously optimizing several
133: contradictory objectives. For such kind of problems, there does not
134: exist a single optimal solution, and compromises have to be made.
135:
136: An element of the search space $x$ is said to {\em Pareto-dominate}
137: another element $y$ if $x$ is not worse than $y$ with respect to all
138: objectives, and is strictly better than $y$ with respect to at least
139: one objective. The set of all elements of the search space that are
140: not Pareto-dominated by any other element is called the {\em
141: Pareto set} of the multi-objective problem at hand: it represents the
142: best possible compromises with respect to the contradictory
143: objectives.
144:
145: Solving a multi-objective problem amounts to choose one solution among
146: those non-dominated solutions, and some decision arguments have to be
147: given. Unlike classical optimization methods, that generally find one
148: of the Pareto optimal solutions by making the initial optimization
149: problem single-objective, EMAs are to-date the only algorithms that
150: directly search for the whole Pareto set, allowing decision makers
151: to choose one of the Pareto solutions with more complete information.
152:
153:
154: \subsection{Pareto-based Evolutionary Algorithms}
155:
156: In order to find a good approximation of the Pareto set (a uniform and
157: well spread sampling of the non-dominated solutions, close
158: to the actual Pareto set of the problem at hand), EMAs have to enforce some
159: progress toward the Pareto set while, at the same time, preserving
160: diversity between the non-dominated solutions.
161:
162: Numerous evolutionary methods have been designed in the past years for
163: the particular task of searching for the Pareto set (the interested
164: reader will find a good summary in~\cite{book}). The best performing
165: among them (NSGA-II~\cite{NSGAII:PPSN2000}, SPEA2~\cite{SPEA2:2001},
166: PESA~\cite{PESA:PPSN2000}) are directly based on the Pareto dominance
167: relation, that actually ensures progressing toward the non-dominated
168: set.
169:
170: Among the diversity preserving techniques, some were transposed to
171: EMAs from single-objective EAs (such as sharing, for instance),
172: while others, like the crowding distance %that will be
173: described in next subsection, are specific to the multi-objective
174: framework.
175:
176: Another recognized important feature of EMAs is
177: elitism~\cite{ZDT:2000}, directly related to the notion of the Pareto
178: dominance in EMAs: the non-dominated individuals can be preserved
179: either by maintaining an archive (SPEA2 and PESA) or by using a
180: deterministic replacement procedure (NSGA-II).
181:
182: \subsection{NSGA-II}
183: \label{NSGA-II}
184:
185: The NSGA-II algorithm has been proposed by Deb et al. in
186: 2001~\cite{NSGAII:PPSN2000}. The progress toward the Pareto set is
187: here favored by using a selection based on the {\em Pareto ranking},
188: that divides the population into a hierarchy of non-dominated subsets,
189: as illustrated by figure~\ref{nsga2fig}(a). All non-dominated
190: individuals of the population are first labeled as being of rank~1;
191: they are then temporarily removed from the population, and the process
192: is repeated: the non-dominated individuals of the remainder of the
193: population are given rank~2, and so on, until the whole population is
194: ranked.
195:
196:
197: \begin{figure}[!h]
198: \begin{center}
199: \begin{tabular}{ccc}
200: \includegraphics[width=3.5truecm]{nsgafit.eps}&\mbox{\hspace{0.cm}}&
201: \includegraphics[width=4.5truecm]{crowdd.eps}\\
202: \\
203: \footnotesize{\bf (a) } Ensuring progress toward the Pareto set & \mbox{\hspace{0.2cm}}&
204: \footnotesize{\bf (b) } Preserving diversity technique
205: \end{tabular}
206: \end{center}
207: \vspace{-0.2cm}
208: \caption{\small\it NSGA-II comparison criteria}
209: \label{nsga2fig}
210: \end{figure}
211:
212: NSGA-II diversity preserving technique is based on the {\it crowding
213: distance} - one of the possible estimations of the density of the
214: solutions belonging to the same non-dominated subset. The crowding
215: distance of each individual $i$ is computed as follows: the
216: non-dominated subset to which the individual $i$ belongs is ordered
217: following each of the objectives; for each objective $m$, the distance
218: $d_i^{(m)}=f_m(i+1)-f_m(i-1)$ between the surrounding neighbors of
219: individual $i$ according to objective $m$ is computed
220: (Fig.~\ref{nsga2fig}(b)); the sum over all objectives of these
221: distances is the crowding distance of individual $i$.
222:
223: The following comparison operator $\succ$ is then used
224: during the Darwinian stages (selection and replacement) of NSGA-II:\\
225:
226: \noindent
227: \begin{tabular}{l}
228: $\quad x \succ y$ {\bf iff} $\quad$rank$(x)$ $<$ rank$(y)$\\
229: \hspace{2cm} {\bf or} $\quad$rank$(x)$ $=$ rank$(y)$\\
230: \hspace{2.9cm} {\bf and } $\quad$crowding\_dist$(x)$ $>$ crowding\_dist$(y)$$\quad$
231: \end{tabular}
232:
233: NSGA-II selection is based on tournament: it chooses
234: an individual for reproduction by uniformly drawing $T$ individuals
235: (generally, $T=2$) from the population and returning the best of them
236: with respect to the comparison operator $\succ$.
237: NSGA-II replacement is deterministic:
238: it consists in merging parents and offspring together,
239: and choosing the $N$ best individuals in that global population,
240: again using comparison operator $\succ$.
241: The algorithm NSGA-II is said to be elitist because
242: the best (less crowded non-dominated) individuals are preserved from
243: one generation to another.
244:
245:
246: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
247: \section{Dominance-based crossover}
248: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
249: \label{xover}
250:
251: The basic mechanism of the proposed mating restriction consists in allowing
252: the mating of each of the best individuals only with an individual it
253: dominates (if any), where {\it best individuals} means
254: non-dominated individuals when applying NSGA-II, or members of the
255: archive when using SPEA2 ou PESA.
256:
257: The rationale behind using the dominance relation to form the couples for
258: the crossover is the following. If $x$ dominates $y$, then $x$ is better
259: than $y$ for all objectives. Hence, the direction $y \rightarrow x$ is
260: likely to improve all criteria simultaneously.
261: Furthermore, a natural continuation of the same idea is to bias the
262: distribution of the offspring toward the dominant parent, as better
263: individuals are more likely to be found close to it.
264: However, it is clear that success of this idea depends
265: on the behavior of the objective functions in the region of the
266: decision space where the mated individuals sit.
267:
268: The resulting crossover, called Dominance-Based Crossover (DBX)
269: will proceed as follows: a first mate is chosen using
270: the usual selection procedure of the EMA at hand (e.g. tournament based
271: on the $\succ$ operator for NSGA-II). If the chosen individual is non-dominated
272: and dominates some other individuals in the population, its mate is chosen among
273: those. Otherwise, the mate is chosen using the usual selection
274: procedure. In any case, crossover then proceeds with the chosen
275: operator.
276:
277: In this study, the well-known BLX-$\alpha$ crossover ($0<\alpha<1$),
278: proposed by Eshelman and Schaffer~\cite{Eshelman:Schaffer:93} for
279: real representations, has been used. Formally, given two
280: parents $(x_i)_{i\in[1,n]}$ and $(y_i)_{i\in[1,n]}$, this operator
281: produces an offspring by a random linear recombination of the parents
282: as follows:
283:
284: \vspace{-0.3cm}
285: \begin{equation}
286: \label{blx}
287: \left( (x_i)_{i\in[1,n]}, (y_i)_{i\in[1,n]} \right) \longrightarrow
288: \left(\phi_i x_i + (1-\phi_i) y_i \right)_{i\in[1,n]},
289: \end{equation}
290: where $\phi_i = U[\alpha, 1+\alpha]$. In our particular case, given a
291: non-dominated indi\-vi\-dual $(x_i)_{i\in[1,n]}$ from the NSGA-II
292: population, two possible strategies will be considered to generate the
293: offspring:
294:
295: \vspace{-0.3cm}
296: \begin{enumerate}
297:
298: \item {\small\bf Symmetric DBX}:
299: The mate $(y_i)_{i\in[1,n]}$ is chosen
300: from the list of the individuals dominated by $(x_i)_{i\in[1,n]}$, if
301: any, by tournament otherwise, and $\phi_i = U[-0.5, 1.5]$ in
302: Equation~(\ref{blx}).
303:
304: \item {\small\bf Biased DBX}:
305: Similarly, the mate $(y_i)_{i\in[1,n]}$ is chosen
306: from the list of the individuals dominated by $(x_i)_{i\in[1,n]}$, if
307: any, but now $\phi_i = U[0.5, 1.5]$ in Equation~(\ref{blx}), i.e. the
308: offspring will be closer to the first parent $(x_i)_{i\in[1,n]}$.
309: \end{enumerate}
310:
311:
312: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
313: \section{Experimental results}
314: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
315: \label{experiments}
316:
317: \subsection{Experimental conditions}
318: This section presents some experimental results, on three standard
319: benchmark problems~\cite{ZDT:2000} and on an
320: industrial problem~\cite{EAmultipla}. All experiments are run with
321: population size 100 and tournament size 2. The two DBX
322: crossovers are compared to the
323: standard BLX-0.5, the crossover rate is set to 0.9.
324: The uniform mutation operator is
325: applied with rate 0.05. The algorithms run for at most
326: 150 (resp. 250) generations for ZDT-benchmarks (resp.
327: the industrial problem).
328:
329: \subsection{Bi-objective ZDT benchmarks}
330:
331: For each of ZDT1--ZDT3 test problems 31 NSGAII runs have been performed
332: for biased DBX, symmetric DBX and standard BLX-0.5 operators
333: starting with the same 31 initial populations.
334: The non-dominated individuals over all 31 runs have been calculated
335: at each 10th generation approximately until the moment when the whole
336: population is non-dominated, that means that DBX crossover
337: is not applied any longer.
338: These snapshots corresponding
339: to biased DBX and standard BLX-0.5 are
340: shown in the figures~\ref{zdt1}, ~\ref{zdt2} and ~\ref{zdt3}
341: for ZDT1, ZDT2 and ZDT3 respectively.
342: %
343: On all three test problems, a small but steady acceleration of the
344: progress toward the Pareto front is observed when using DBX crossover.
345: On those problems, very similar results have been obtained with
346: the symmetric and biased DBX operators (the snapshots corresponding
347: to the symmetric DBX are not shown for the space reasons).
348: %
349:
350: \begin{figure}[!h]
351: \vspace{-0.2cm}
352: \begin{center}
353: \begin{tabular}{cc}
354: %\includegraphics[width=5truecm]{POPS/1_0.ps}&\includegraphics[width=5truecm]{POPS/1_10.ps}\\
355: \includegraphics[width=5truecm]{POPS/1_10.ps}&\includegraphics[width=5truecm]{POPS/1_20.ps}\\
356: \includegraphics[width=5truecm]{POPS/1_30.ps}&\includegraphics[width=5truecm]{POPS/1_40.ps}\\
357: \end{tabular}
358: \end{center}
359: \vspace{-0.2cm}
360: \caption{\footnotesize\it {\bf ZDT1}: the black bullets (for the biased DBX)
361: and the gray crosses (for the standard BLX-0.5) represent the non-dominated
362: individuals over 31 runs at generations 10, 20, 30 and 40.}
363: \label{zdt1}
364: \end{figure}
365: %
366:
367: \vspace{0.1cm}
368: \begin{figure}[!h]
369: \begin{center}
370: \begin{tabular}{cc}
371: \includegraphics[width=5truecm]{POPS/2_20.ps}&\includegraphics[width=5truecm]{POPS/2_30.ps}\\
372: \includegraphics[width=5truecm]{POPS/2_40.ps}&\includegraphics[width=5truecm]{POPS/2_50.ps}\\
373: \end{tabular}
374: \end{center}
375: \vspace{-0.2cm}
376: \caption{\small\it {\bf ZDT2}: the black bullets (for the biased DBX)
377: and the gray crosses (for the standard BLX-0.5) represent the non-dominated
378: individuals over 31 runs at generations 20, 30, 40 and 50.}
379: \label{zdt2}
380: \end{figure}
381: %
382:
383: \begin{figure}[!h]
384: \vspace{0.2cm}
385: \begin{center}
386: \begin{tabular}{cc}
387: \includegraphics[width=5truecm]{POPS/3_10.ps}&\includegraphics[width=5truecm]{POPS/3_20.ps}\\
388: \includegraphics[width=5truecm]{POPS/3_30.ps}&\includegraphics[width=5truecm]{POPS/3_40.ps}\\
389: \end{tabular}
390: \end{center}
391: \vspace{-0.2cm}
392: \caption{\small\it {\bf ZDT3}: the black bullets (for the biased DBX)
393: and the gray crosses (for the standard BLX-0.5) represent the non-dominated
394: individuals over 31 runs at generations 10, 20, 30 and 40.}
395: \label{zdt3}
396: \end{figure}
397:
398: \subsection{Constrained bi-objective optimization}
399:
400: This industrial problem consists in optimizing the structural
401: geometry, described by 13 continuous parameters, of the front crash
402: member of the car in order to minimize its mass while maximizing the
403: internal energy absorbed during a crash (two competitive objectives)
404: under 8 constraints arising from the acoustic and static mechanical
405: domains~\cite{EAmultipla}.
406:
407:
408: The constraints have been handled using the
409: so-called {\it infeasibility objective
410: approach}~\cite{WrightLoosemore:EMO1}:
411: the aggregated sum of the scaled constraint violations (that can be
412: viewed as a measure of distance separating each individual from the
413: feasible region) was considered as an additional optimization
414: criterion -- the infeasibility objective.
415: NSGA-II was hence applied
416: to a three-objective optimization problem, with the difference that
417: the infeasibility objective had a higher priority than both others in
418: the $\succ$ operator.
419: In the other words, every individual located closer to the feasible
420: region is considered better than any individual further away
421: from it, regardless the corresponding values of the other objectives,
422: the mass and the absorbed internal energy.
423:
424: \begin{figure}[!h]
425: \vspace{-0.2cm}
426: \begin{center}
427: \begin{tabular}{cc}
428: \includegraphics[width=6truecm]{constr_a.ps}&\includegraphics[width=6truecm]{constr_b.ps}\\
429: {\small\bf (a)} \tiny Faster reaching the feasible region &
430: {\small\bf (b)} \tiny Better sampling the Pareto front extremities
431: \end{tabular}
432: \end{center}
433: \vspace{-0.2cm}
434: \caption{\small\it Effect of the dominance-based mating restriction
435: in the presence of constraints:
436: biased DBX (rings) VS standard BLX-0.5 (crosses)}
437: \label{constraints}
438: \end{figure}
439:
440: DBX operators can be employed in this context exactly like
441: for unconstrained problems, using the modified $\succ$ comparison.
442: On every run, the restricted mating allowed faster approach of the
443: feasible region, as well as faster feasibility of the whole population,
444: as illustrated by Figure~\ref{constraints}-a. Moreover, some
445: differences with the results on the test problems ZDT have
446: been steadily observed. First, biased DBX was more efficient than
447: symmetric DBX. Then, and more importantly, the use of mating
448: restriction not only accelerated the search, it also provided
449: solutions of better quality at the extremities of the Pareto front,
450: as it can be seen on figure~\ref{constraints}-b.
451:
452: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
453: \section{Discussion and future work}
454: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
455: \label{conclusion}
456:
457: For all four problems considered in this study, the DBX
458: operators (that only allow the mating of
459: dominant individuals with individuals they dominate)
460: have been shown to accelerate the progress
461: of the populations toward the Pareto set. Moreover,
462: for the optimization of the car front crash member, it also allowed
463: finding solutions of better quality at the
464: extremities of the Pareto set that could not be reached when using
465: the usual recombination strategy.
466:
467: %
468: \begin{figure}[!h]
469: \begin{center}
470: \begin{tabular}{cc}
471: \includegraphics[width=5truecm]{concl_a.ps}&
472: \includegraphics[width=5truecm]{concl_b.ps}\\
473: {\bf\small(a)} \tiny Number of non-dominated individuals&
474: {\bf\small(b)} \tiny 12 non-dominated inds at the 10th gen. \\
475: & \\
476: \includegraphics[width=5truecm]{concl_c.ps}&
477: \includegraphics[width=5truecm]{concl_d.ps}\\
478: {\bf\small (c)} \tiny 47 non-dominated inds at the 35th gen.&
479: {\bf\small (d)} \tiny 100 non-dominated inds at the 43rd gen.
480: \end{tabular}
481: \end{center}
482: \vspace{-0.2cm}
483: \caption{\footnotesize\it Evolution of the number of non-dominated
484: individuals during a run of NSGA-II}
485: \label{nondNb}
486: \vspace{-0.5cm}
487: \end{figure}
488: %
489: The observed acceleration is relatively small, but is
490: systematic. When observing NSGA-II dynamics (the typical NSGA-II
491: evolution process is illustrated by Figure~\ref{nondNb}), we realize that
492: in fact only a small number of DBX crossovers are actually applied.
493: First, note that restricted mating is applied at most as many
494: times as there are non-dominated individuals in the population.
495: There are few non-dominated individuals at the beginning of evolution
496: (Fig.~\ref{nondNb}-a,b), but each of them dominates
497: a lot of other individuals (Fig.~\ref{nondNb}-b).
498: As the population gets closer to the Pareto set, the number of
499: non-dominated individuals rapidly increases (Fig.~\ref{nondNb}-a,c),
500: but each of them dominates only a few individuals if any (Fig.~\ref{nondNb}-c).
501: Finally, when the whole
502: population is non-dominated (Fig.~\ref{nondNb}-d),
503: DBX crossover cannot be applied because
504: no individual actually dominates anyone in the population!
505: The actual rate of application of the DBX operator
506: is maximal about the moment when the half of the
507: population gets the rank~1 (Fig.~\ref{nondNb}-c).
508:
509: One possible improvement would be to increase the rate of
510: application of DBX operators at the beginning of evolution
511: by mating each non-dominated individual with not only one but several
512: individuals it dominates.
513:
514: Note that dynamic behavior of the populations described here above
515: and, in particular,
516: the disappearance of the dominated individuals is due to the
517: the replacement procedure in NSGA-II, and might not be so
518: prominent with other EMAs.
519: It is hence worth investigating the use of DBX restricted mating
520: with other EMAs paradigms, such as SPEA2 or PESA, for instance.
521:
522: Furthermore, there exist other crossover operators used in EAs (such
523: as SBX~\cite{book}, for example) that
524: could be applied together with dominance based restricted mating,
525: instead of BLX-$\alpha$.
526:
527: % ----------------------------------------------------------------------
528: One more issue, that needs to be thoroughly investigated, is the
529: efficiency of the DBX strategy when solving problems with more than
530: two objectives. Indeed, in such situations, the
531: ``visual'' analysis of the populations dynamics performed in the present work
532: will not be possible anymore, and our plans include the use of so-called
533: {\it running metrics}~\cite{DebJain:2002}.
534:
535: Indeed, the present study could use
536: some performance metrics as well, instead of the graphical plots of
537: the population dynamics in the objective space.
538: However, in our opinion, such presentation is much clearer and makes
539: things more explicit, hence it is more appropriate to the introduction
540: of the new operator.
541: %----------------------------------------------------------------------
542:
543:
544: But probably the most important issue is to find a way to actually
545: benefit from the reported acceleration in practice. Whereas its
546: usefulness is obvious for costly applications, where the number of
547: evaluations has to be limited, the question remains of how to detect
548: that the algorithm is getting stuck, thus saving the computational
549: time that restricted mating allows us to spare.
550: An efficient stopping criterion is needed before any actual benefit
551: can be gained from the observed acceleration.
552: Such criterion has been proposed by the first author
553: \cite{these-Olga,MOPGP}, but it was beyond the scope of this paper.
554: Nevertheless, together with the DBX operators described in this paper, it
555: allowed an actual saving of about 8\% of the computation time.
556: However, further evaluation and refinement of that stopping criterion
557: are still needed before definite strong conclusions can be drawn.
558:
559:
560: %----------------------------
561: %\bibliography{../laBiblio}
562:
563: \begin{thebibliography}{10}
564:
565: \bibitem{PESA:PPSN2000}
566: D.~W. Corne, J.~D. Knowles, and M.L. Oates.
567: \newblock The Pareto Envelope-based Selection Algorithm for Multiobjective
568: Optimization.
569: \newblock In M.~Schoenauer and al., editors, {\em Proc. PPSN 2000}, pages 839--848. LNCS 1917, Springer Verlag, 2000.
570:
571: \bibitem{book}
572: K.~Deb.
573: \newblock {\em Multi-Objective Optimization using Evolutionary Algorithms}.
574: \newblock John Wiley, 2001.
575:
576: \bibitem{NSGAII:PPSN2000}
577: K.~Deb, S.~Agrawal, A.~Pratap, and T.~Meyarivan.
578: \newblock A Fast Elitist Non-dominated Sorting Genetic Algorithm for
579: Multi-objective Optimization: {NSGA-II}.
580: \newblock In M.~Schoenauer and al., editors, {\em Proc. PPSN 2000},
581: pages 849--858. LNCS 1917, Springer Verlag, 2000.
582:
583: \bibitem{DebJain:2002}
584: K.~Deb and S.~Jain.
585: \newblock Running performance metrics for evolutionary multi-objective
586: optimization.
587: \newblock Technical Report 2002004, Indian Institute of Technology, Kanpur, May
588: 2002.
589:
590: \bibitem{Eshelman:Schaffer:93}
591: L.~Eshelman and J.~D. Schaffer.
592: \newblock Real-coded genetic algorithms and interval-schemata.
593: \newblock In L.~D. Whitley, editor, {\em Foundations of Genetic Algorithms 2},
594: pages 187--202, Morgan Kaufmann, 1993.
595:
596: \bibitem{FF93}
597: Carlos~M. Fonseca and Peter~J. Fleming.
598: \newblock Genetic algorithms for multiobjective optimization: Formulation,
599: discussion and generalization.
600: \newblock In {\em ICGA'93}, pages 416--423. Morgan Kaufmann, 1993.
601:
602: \bibitem{Goldberg:89}
603: D.~E. Goldberg.
604: \newblock {\em Genetic algorithms in search, optimization and machine
605: learning}.
606: \newblock Addison Wesley, 1989.
607:
608: \bibitem{HajelaLin:92}
609: P.~Hajela and C.~Y. Lin.
610: \newblock Genetic search strategies in multicriterion optimal design.
611: \newblock {\em Structural Optimization}, 4:99--107, 1992.
612:
613: \bibitem{Horn:al:94}
614: J.~Horn, S.~N. Nafpliotis, and D.~E. Goldberg.
615: \newblock A niched pareto genetic algorithm for multiobjective optimization.
616: \newblock In Z.~Michalewicz et al., editors, {\em Proc. of ICEC'94},
617: pages 82--87. IEEE Press, 1994.
618:
619: \bibitem{these-Olga}
620: O.~Rudenko.
621: \newblock {\em Application des algorithmes évolutionnaires aux problèmes
622: d'opti\-mi\-sa\-tion multi-objectif avec contraintes}.
623: \newblock PhD thesis, École Polytechnique, 2004.
624:
625: \bibitem{MOPGP}
626: O.~Rudenko and M.~Schoenauer.
627: \newblock A steady performance stopping criterion for pareto-based evolutionary
628: algorithms.
629: \newblock In {\em Proc. 6th Intl Conf. on Multi
630: Objective Programming and Goal Programming}, 2004.
631:
632: \bibitem{EAmultipla}
633: O.~Rudenko, M.~Schoenauer, T.~Bosio, and R.~Fontana.
634: \newblock A multiobjective evolutionary algorithm for car front end design.
635: \newblock In P.~Collet et al.,
636: editors, {\em Artificial Evolution'2001}, pages 205--216,
637: LNCS 2310, Springer Verlag, 2001.
638:
639: \bibitem{WildmanParks:EMO2}
640: A.~Wildman and G.~Parks.
641: \newblock A comparative study of selective breeding strategies in a
642: multiobjective genetic algorithm.
643: \newblock In C.~M. Fonseca et al., editors,
644: {\em Proc. EMO'03}, pages 418--432. LNCS 2632, Springer Verlag, 2003.
645:
646: \bibitem{WrightLoosemore:EMO1}
647: J.~Wright and H.~Loosemore.
648: \newblock An infeasibility objective for use in constrained pareto
649: optimization.
650: \newblock In E. Zitzler et al., editors, {\em Proc. EMO'01},
651: pages 256--268, LNCS 1993, Springer-Verlag, 2001.
652:
653: \bibitem{ZDT:2000}
654: E.~Zitzler, K.~Deb, and L.~Thiele.
655: \newblock Comparison of multiobjective evolutionary algorithms: Empirical
656: results.
657: \newblock {\em Evolutionary Computation}, 8(2):125--148, 2000.
658:
659: \bibitem{SPEA2:2001}
660: E.~Zitzler, M.~Laumanns, and L.~Thiele.
661: \newblock {SPEA2}: Improving the strength pareto evolutionary algorithm.
662: \newblock Technical Report 103, Computer Engineering and Networks Laboratory,
663: ETH, Zurich, Switzerland, 2001.
664:
665: \bibitem{ZitzlerThiele:PPSN98}
666: E.~Zitzler and L.~Thiele.
667: \newblock Multiobjective optimization using evolutionary algorithms - a
668: comparative case study.
669: \newblock In A.-E. Eiben, T.~B{\"a}ck, M.~Schoenauer, and H.-P. Schwefel,
670: editors, {\em Proc. PPSN'98}, pages 292--301. LNCS 1498, Springer Verlag, 1998.
671:
672: \end{thebibliography}
673:
674:
675: \end{document}
676:
677: