Controlling Genetic Algorithms With Reinforcement Learning
James E. Pettinger, Richard Everson · 2002
Here we present a hybrid system that uses a reinforcement learning agent to improve the performance of a genetic algorithm on the travelling salesman problem (TSP). The agent uses Q(λ) learning to estimate state-action utility values, which it uses to implement high-level adaptive control over the genetic algorithm. In this way the agent influences selection of both crossover and mutation operators as well as the selection of individuals for breeding, at every generation.