A hybrid ACO algorithm based on Bayesian factorizations and reinforcement learning for continuous optimization

摘要

Ant colony optimization (ACO) is an evolutionary computing approach for combinatorial optimization problems. Recently, some extensions of ACO have been proposed in continuous domains. However, these methods did not consider the dependency between variables and thus may fail for some complex optimization problems. In this paper, we use Bayesian factorizations to capture the main dependency of the variables and sample the more reasonable solutions from the probabilistic models obtained. Inspired by multiagent consensus protocols, we use the neighborhood information of the new solution generated by each ant to enhance its local search ability. However, instead of using all of the neighbors of that solution, we try to distinguish the neighbor which can optimize the performance of the ant from all of the neighbors. Given this situation, reinforcement learning is used to determine the optimal strategy for each ant in the iteration by maximizing both the immediate reward and the delayed reward. The proposed algorithm is compared with some other existing continuous ACO algorithms. Experiments indicate that the proposed algorithm clearly outperforms the other methods investigated and can greatly improve the rate of convergence.