Forming Neural Networks through Efficient and Adaptive Coevolution：（形成神经网络通过高效和适应的共同进化）.pdfVIP

Evolutionary Computation In Press Forming Neural Networks through Ecient and Adaptive Co evolution  David E Moriarty Risto Miikkulainen Information Sciences Institute Department of Computer Sciences University of Southern California The University of Texas at Austin Admiralty Way Austin TX Marina del Rey CA ristocsutexasedu moriartyisiedu Abstract This article demonstrates the advantages of a co op erative co evolutionary search in dicult control problems The SANE system co evolves a p opulation of neurons that co op erate to form a functioning neural network In this pro cess neurons assume dierent but overlapping roles resulting in a robust enco ding of control b ehavior SANE is shown to b e more ecient more adaptive and maintain higher levels of diversity than the more common networkbased p opula tion approaches Further empirical studies illustrate the emergent neuron sp ecializations and the dierent roles the neurons assume in the p opulation Intro duction Articial evolution has b ecome an increasingly p opular metho d for forming control p olicies in dicult decision problems Grefenstette Ramsey Schultz Moriarty Miikkulainen a Whitley Dominic Das Anderson Such applications are very dierent from the function optimization tasks to which evolutionary algorithms EA have b een traditionally applied For exam