Adaptive Reinforcement-Based Genetic Algorithm for Combinatorial Optimization

Combinatorial optimization is the procedure of optimizing an objective function over the discrete configuration space. A genetic algorithm (GA) has been applied successfully to solve various NP-complete combinatorial optimization problems. One of the most challenging problems in applying GA is selecting mutation operators and associated probabilities for each situation. GA uses just one type of mutation operator with a specified probability in the basic form. The mutation operator is often selected randomly in improved GAs that leverage several mutation operators. While an e ective GA search occurs when the mutation type for each chromosome is selected according to mutant genes and the problem landscape. This paper proposes an adaptive genetic algorithm that uses Q-learning to learn the best mutation strategy for each chromosome. In the proposed method, the success history of the mutant in solving the problem is utilized for specifying the best mutation type. For evaluating adaptive genetic algorithm, we adopted the traveling salesman problem (TSP) as a well-known problem in the field of optimization. The results of the adaptive genetic algorithm on five datasets show that this algorithm performs better than single mutation GAs up to 14% for average cases. It is also indicated that the proposed algorithm converges faster than single mutation GAs.