Reduction of computational cost in optimization of parameter values in reinforcement learning by a genetic algorithm

Reinforcement learning (RL), suitable for navigation of a mobile robot, has a difficulty where parameter values can only be determined by trial and error. We proposed to use a genetic algorithm (GA) with inheritance to obtain optimal parameter values in RL, which reduced the computational cost by 99% compared with that without inheritance. Since the computational cost is still 2 orders of magnitude larger than that by the conventional RL, we propose further reduction by decreasing the number of generations, the number of individuals, and the number of episodes. We succeed in further reducing the computational cost by about 75% compared with our previous proposal of RL using GA with inheritance.