Critical dynamics in evolutionary algorithms

Genetic algorithms (GA) have proved to be an effective technique for search and optimization over difficult domains. One common problem for GAs is the phenomenon of premature convergence to suboptimal solutions. We conjecture that premature convergence occurs in part because genetic algorithms lack critical dynamics. This paper proposes a novel algorithm, the genepile evolutionary algorithm, which makes use of the complex spatial dynamics of the sandpile model of self-organized criticality. It is suggested that the critical dynamics of this algorithm make it less prone to getting trapped at local optima. Though the genepile evolutionary algorithm did converge during testing, it has nonetheless proved to be an effective optimization tool, recording good performance across a broad suite of test functions and in many cases substantially outperforming two well-known control algorithms.