Multi-Agent Systems of Inverse Reinforcement Learners in Complex Games

Real-world problems exhibit a few defining criteria that make them hard for computers to solve. Problems such as driving a car or flying a helicopter have primary goals of reaching a destination as well as doing it safely and timely. These problems must each manage many resources and tasks to achieve their primary goals. The tasks themselves are made up of states that are represented by variables or features. As the feature set grows, the problems become intractable. Computer games are smaller problems but also are representative of real-world problems of this type. In my research, I will look at a particular class of computer game, namely computer role-playing games (RPGs), which are made up of a collection of overarching goals such as improving the player avatar, navigating a virtual world, and keeping the avatar alive. While playing there are also subtasks such as combatting other characters and managing inventory which are not primary, but yet important to overall game play. I will be exploring tiered Reinforcement Learning techniques coupled with training from expert policies using Inverse Reinforcement Learning as a starting point on learning how to play a complex game while attempting to extrapolate ideal goals and rewards.