Year of Graduation
Reinforcement Learning in CCG-like (Collectible Card Game) Environments
Applied Mathematics and Information Science
Reinforcement learning researches in most cases based on different environments and their features. For example, Atari games like “Montezuma Revenge” and “Pitfall” pose the “sparse rewards” problem and as the result methods like Go-Explore was invented. Other Atari games allowed creating a “Rainbow” method. More complex game-environments like DotA 2 or Starcraft 2 pose new non-solved problems. Based on this approach, I consider it necessary to study CCG-like environments (collectible card games like TES Legends, Hearthstone, Eternal, etc), especially because environments like this never was examined carefully. This type of environments contains an interesting combination of features interesting in the scope of the reinforcement learning task. In this paper, this environment has been implemented and investigated, some inner environmental problems have been identified and solved. High results were achieved in various environments using such algorithms as a2c, self-play and their modifications.