Evolutionary methods are powerful tools in discovering solutions for difficult continuous tasks. When such a solution is encoded over multiple genes, a genetic algorithm faces the difficult credit assignment problem of evaluating how a single gene in a chromosome contributes to the full solution. Typically a single evaluation function is used for the entire chromosome, implicitly giving each gene in the chromosome the same evaluation. This method is inefficient because a gene will get credit for the contribution of all the other genes as well. Accurately measuring the fitness of individual genes in such a large search space requires many trials. This paper instead proposes turning this single complex search problem into a multi-agent search problem, where each agent has the simpler task of discovering a suitable gene. Gene-specific evaluation functions can then be created that have better theoretical properties than a single evaluation function over all genes. This method is tested in the difficult double-pole balancing problem, showing that agents using gene-specific evaluation functions can create a successful control policy in 20% fewer trials than the best existing genetic algorithms. The method is extended to more distributed problems, achieving 95% performance gains over tradition methods in the multi-rover domain.

PDF

In Proceedings of the Genetic and Evolutionary Computation Conference, 2005.

@incollection{agogino:gecco05, title={Efficient Credit Assignment through Evaluation Function Decomposition}, author={Adrian Agogino and Kagan Tumer and Risto Miikkulainen}, booktitle={Proceedings of the Genetic and Evolutionary Computation Conference}, month={ }, url="http://nn.cs.utexas.edu/?agogino:gecco05", year={2005} }