Learning from Stories: Using Natural Communication to Train Believable Agents

In this work we introduce Quixote, a system that allows non-programmers to train believable virtual agents and robots using the sociocultural knowledge present in stories. Quixote uses a corpus of exemplar stories to engineer a reward function that can be used to train virtual agents to exhibit desired behaviors using reinforcement learning. We show the effectiveness of our system with a case study in a modified gridworld environment called Pharmacy World. In this case study, we examine the performance of Quixote under ideal conditions and evaluate how humans perceive the policies that our system produces. In addition, we examine how Quixote performs in this environment when ambiguities in natural language correspondence causes difficulty in mapping story events to agent actions.