Efficient Credit Assignment through Evaluation Function Decomposition

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 percent fewer trials than the best existing genetic algorithms. The method is extended to more distributed problems, achieving 95 percent performance gains over tradition methods in the multi-rover domain.