Competitive Deep Reinforcement Learning over a Pokémon Battling Simulator

摘要

Pokémon is one of the most popular video games in the world, and recent interest has appeared in Pokémon battling as a testbed for AI challenges. This is due to Pokémon battling showing interesting properties which contrast with current AI challenges over other video games. To this end, we implement a Pokémon Battle Environment, which preserves many of the core elements of Pokémon battling, and allows researchers to test isolated learning objectives. Our approach focuses on type advantage in Pokémon battles and on the advantages of delayed rewards through switching, which is considered core strategies for any Pokémon battle. As a competitive multi-agent environment, it has a partially-observable, high-dimensional, and continuous state-space, adheres to the Gym de facto standard reinforcement learning interface, and is performance-oriented, achieving thousands of interactions per second in commodity hardware. We determine whether deep competitive reinforcement learning algorithms, WPLθ and GIGAθ, can learn successful policies in this environment. Both converge to rational and effective strategies, and GIGAθ shows faster convergence, obtaining a 100% win-rate in a disadvantageous test scenario.