Automatic learning and generation of social behavior from collective human gameplay

Abstract

Current approaches to authoring behavior and dialogue for agents that interact with humans in virtual environments are labor intensive, yet often yield less robust results than desired in the face of the incredible variance possible in human input. The growing number of people playing multiplayer games online provides a potentially better alternative to hand-authored content -- capturing behavior and dialogue from human-human interactions, and automating agents with this data. This paper documents promising results from the first iteration of a Collective Artificial Intelligence system that generates behavior and dialogue in real-time from data captured from over 11,000 players of The Restaurant Game. We first describe the game, the collective memory system, and the proposal-critique driven agent architecture, and then demonstrate quantitatively that our system preserves the texture, or meaningful local coherence, of human social interaction.