Agents with incomplete models of their environment are likely to be surprised by it. For agents in immense environments that defy complete modeling, this represents an opportunity to learn. We investigate approaches for situated agents to detect surprise, discriminate among different forms of surprise, and ultimately hypothesize new models for the unknown events that surprised them. We instantiate these approaches in a new goal reasoning agent, FooLMETWICE, and investigate how that agent performs in a simulated environment. In this case study, we found that FooLMETWicE learn models that substantially improve its performance.
展开▼