Parallel Recombinative Reinforcement Learning

A technique is presented which is suitable for function optimization in high-dimensional binary domains. The method allows an eecient parallel implementation and is based on the combination of genetic algorithms and reinforcement learning schemes. More speciically, a population of probability vectors is considered, each member corresponding to a reinforcement learning optimizer. Each probability vector represents the adaptable parameters of a team of stochastic units whose binary outputs provide a point of the function state space. At each step of the proposed technique the population members are updated according to a reinforcement learning rule and then recombined in a manner analogous to traditional genetic algorithm operation. Special care is devoted to ensuring the desirable properties of sustained exploration capability and sustained population diversity. The eeciency of the method is tested on two deceptive problems that constitute typical benchmarks yielding very promising results.