Learning Game Design

Abstract

Action video games feature many game variables such as monster health and dam- age which control the difficulty of the game. To help automate the search for parameters which provide reasonable challenge to players, we present ”Dungeon Master,” a Reinforcement Learning agent which uses Continuous Actor Critic to find a game parameters that meet some high level, designer specified difficulty. We implemented a minimal, turn-based action game featuring rooms of monsters whose game statistics are controlled by the Dungeon Master. We created a novel variable update method, which allows our agent to step game variables towards a configuration that meets a difficulty specification, measured in player deaths per room. For a 6 room “dungeon” in our game, Dungeon Master performs as well as a human designer in finding game parameters that meet a specified difficulty for AI player model.