Researches On Improved Genetic Algorithm Base On Reinforcement Learning

Genetic algorithm, which is the earliest, most influential and widely applied, is a branch of Evolutionary Algorithms. As the global optimization and probability searching method, it has many advantages such as simplicity, robustness and commonality. Since it was proposed, it has been applied to engineering, machine learning, pattern recognition, image processing and other aspects. Despite all this, there are many problems should be perfect urgently in theory and application methods, such as premature convergence and slow convergence.In this thesis, some improved methods are proposed to avoid the premature convergence and low speed of convergence. The main research results are concluded as follows:i. A new multiple policy selection genetic algorithm based on reinforcement learning is proposed. The algorithm improves population diversity by using several different selection policies, which can avoid the premature convergence effectively. And population diversity is adjusted by reinforcement learning, which can maintain the population diversity in the appropriate range and enhance the self-adaptive capability to some extent. The experiments show that the algorithm has a high performance in the speed of convergence and search efficiency.ii. Based on the elite strategy and the concept of coevolution, a new double elite coevolutionary genetic algorithm is proposed. And the convergence of the algorithm is proved theoretically. Theoretical analysis proves that the algorithm converge to the global optimization solution. Tests on the functions show that the algorithm can find the global optimal solution for the most test functions, and it can also maintain the population diversity to a certain range. Compared with the existing algorithms, the double elite coevolutionary genetic algorithm has a higher performance in the speed of convergence and search efficiency.iii. In order to improve the self-evolution mechanism of double elite coevolutionary genetic algorithm further and enhance the reliability and self-adaptive capability of the algorithm, reinforcement learning algorithm is applied into double elite coevolutionary genetic algorithm. The experiments show that the combination between reinforcement learning algorithm and genetic algorithm can effectively overcome the randomness of genetic algorithm itself. And it provides a guideline of the combination between genetic algorithm and other reinforcement learning methods.