Reinforcement learning for partially observable stochastic games

Abstract

Inscryption is a single player rogue‐like deck builder game. The aim of this project is to implement the ’card battle’ aspect of the game and then use reinforcement learning algorithms to teach the model how to play. This game deals with a multitude of factors that affect the battle, the ones relevant to this project are the lack of deck information; the order of the deck is not known, the randomness of the battles. The game also contains 57 base cards, each which have sigils that can effect how they attack and are played. This leads to a very large action space of 42,350 possible actions. Another challenge is the issue of rewards, since the goal of winning the game can be achieved in many different way and in some cases the player needs to ’lose’ in the current state in order to win the whole game. The leaves the question on how to represent the rewards. Research was done on how similar project tackled similar games, as no other project that is publicly available tackled this game.