Reinforcement Learning in Interactive Fiction Games
Interactive Fiction (IF) games are a challenging domain combining sequential decision making and natural language. IF games consist of text-based state and action spaces and contain a variety of new problems for learning agents including commonsense reasoning, textual SLAM, and natural-language action spaces. In this talk I'll present initial progress towards solving IF games, in the form of a reinforcement learning agent that plays IF games as well as a heuristic agent designed for human-made games. Finally, I'll discuss the recent Text-Based Adventure competition, where our agent took first place.
Matthew Hausknecht is a researcher at Microsoft Research. He obtained a PhD degree in Computer Science from the University of Texas at Austin. His main research interests are in reinforcement learning and decision making in complex environments. To drive the development of reinforcement learning agents, he has helped create learning environments such as Arcade Learning Environment, Half Field Offense for simulated multi-agent soccer, and Jericho for text-based game playing.