Games are a major testing ground for Artificial Intelligence. Though AI has be- come proficient at playing games such as Space Invaders, it behaves in a way that is distinctly artificial, lacking the human-like qualities of a real player. This human ele- ment is important in competitive multiplayer games, as a large part of the enjoyment comes from outwitting other human strategies. To address this issue, we investigate a novel AI technique that leverages planning and human demonstrations to create an opponent that exhibits desirable qualities of human play in the context of a fight- ing game. We introduce the idea of action-δs, which relate the action performed with the change in the game state. These action-δs are learned from human demon- strations and are used to help the AI plan out strategies to hit the opponent. We implement a simple fighting game called FG for the AI to compete in and provide it a human demonstration to learn from. The AI utilizes action-δs with other search techniques to emulate human behavior. Lastly, we evaluate the effectiveness of our AI by comparing its similarity score against other algorithms and other demonstra- tions by the same human player.

Tech Report: CMU-CS-17-128