Research on Atari Games using Evolutionary Computation with Successive Halving
Takuki Kurokawa, Hitoshi Iba · 2023
Currently, distributed deep reinforcement learning is one of the most popular approaches in game AI research. However, research on game AI using distributed evolutionary computation is limited. These methods utilize large amount of computational resources. Therefore, reducing the computational complexity of the algorithm is very important to improve the training efficiency of game AI agents. In this study, we propose a method for efficiently selecting top individuals from a set of game AI agents, based on Successive Halving. We applied Successive Halving to Genetic Algorithm and trained game AI agents on three Atari 2600 games. Consequently, we succeeded in improving the training efficiency of these three games in the early stages of training. We have concluded that proposed method can be utilized as a fast initial solution generation algorithm. We suggest that hyperparameter optimization techniques can be applied to algorithmic improvements in evolutionary computing.