Go with the Flow: Reinforcement Learning in Turn-based Battle Video Games

Abstract

Game flow represents a state where the player is neither frustrated nor bored. In turn-based battle video games it can be achieved by Dynamic Difficulty Adjustment (DDA), whose research has begun rising over the last decade. This paper introduces an idea for incorporating DDA through the use of Reinforcement Learning (RL) to agents of turn-based battle video games. We design and implement an RL agent that shows, in a simple environment, the idea of how a game could achieve balance through adequate choices in actions depending on the player's level of skill. For achieving this purpose, we incorporated the design and implementation of state-action-reward-state-action (SARSA) algorithm to the agent of our implemented game. In addition, we added tracking of the on-going games and depending on the frequency of the player's repeated wins or losses, the rewards of the RL agent are modified. This modification of the rewards has an impact on the RL agent's actions, which involves an increase/decrease of the difficulty of the battle game. The evaluation performed shows that the idea of the paper is demonstrated, since players face personalized challenges that we believe are in range of game flow.