A Self-Play Policy Optimization Approach to Battling Pokémon
Dan Hong Huang, Scott Lee · 2019 IEEE Conference on Games (CoG) · 2019
Pokémon is a popular role-playing video game franchise with a long-lived competitive scene that has evolved throughout the last two decades. The game exhibits several properties that come together to present a worthy challenge for AI agents to tackle. In this work, we present a low-cost self-play based reinforcement learning approach to the competitive battling aspect of the game. The proposed agent was tested and trained in a variety of environments designed to simulate possible use cases of such an AI. Experiments demonstrate that the agent is capable of performing on par with the state of the art in search-based Pokémon AI, as well as being competitive with́ human players on a popular matchmaking ladder. Furthermore, we investigate the transferability of trained skill-whether an agent trained in one environment performs well in a different environment.