Extraction of Interaction Events for Learning Reasonable Behavior in an Open-World Survival Game

Extracting event knowledge from open-world survival video games is a promising domain to investigate the application of Machine Learning techniques to routine human decision making. This contrasts with and builds upon typical game-based decision making work that focuses on optimal behavior. We propose an Interaction Graph data structure that can be trained from game play to enable hybrid reasoning and statistical estimation about what events can happen in the world. This enables an agent to exhibit increasingly more reasonable behavior after low numbers of training runs. An implementation and initial experimental validation are presented.