SCGA: Controlling Genetic Algorithms with Sarsa(0)

Though deeply analyzing and comparing the mechanism of genetic algorithm and reinforcement learning, a novel algorithm for controlling genetic algorithms using reinforcement learning named SCGA, is proposed and analyzed theoretically. In the existing similar method RL-GA, a reinforcement learning agent uses Q(lambda)-learning to control genetic algorithms. Two problems with such method are that, (1) Q(lambda)-learning cannot fit the stochastic and dynamic characters of genetic algorithms well, and (2) dividing the whole algorithm running time into training and testing schemes reduces its practicability. To solve these drawbacks, SCGA implements the on-policy method Sarsa(0) to control both of the GA operators to choose and the individuals to select without training first. The experimental results show that SCGA learns much faster than RL-GA and the primitive GA on the traveling salesman problem, and is more practicable and scalable in applications