1: \begin{abstract}

2: \noindent

3: Neuroevolution has yet to scale up to complex reinforcement learning

4: tasks that require large networks.  Networks with many inputs (e.g.\

5: raw video) imply a very high dimensional search space if encoded

6: directly.  Indirect methods use a more compact genotype representation

7: that is transformed into networks of potentially arbitrary size.  In this

8: paper, we present an indirect method where networks are encoded by

9: a set of Fourier coefficients which are transformed into network

10: weight matrices via an inverse Fourier-type transform.  Because there

11: often exist network solutions whose weight matrices contain regularity

12: (i.e.\ adjacent weights are correlated), the number of coefficients

13: required to represent these networks in the frequency domain is much

14: smaller than the number of weights (in the same way that natural

15: images can be compressed by ignore high-frequency components).  

16: This ``compressed'' encoding is compared to the direct approach where

17: search is conducted in the weight space on the high-dimensional

18: octopus arm task.  The results show that representing networks in the

19: frequency domain can reduce the search-space dimensionality by as

20: much as two orders of magnitude, both accelerating convergence and

21: yielding more general solutions. 

22: \end{abstract}

23: