Learning Avatar Behaviours Project for MAS 732

To engage in discourse, a participant must map the discourse functions he or she wants to achieve into surface level realizations. These mappings are often treated as rule-based and general in work on discourse and computational discourse. Here, I demonstrate a proof-of-concept system that learns such mappings stochastically in a reinforcement learning paradigm. Learning these mapping has the advantage that it takes some of the work out of the mapping design and allows for more personalized systems. Approaching the problem stochastically deals with the inherent brittleness of rulebased system and allows reasoning about beliefs as well as graceful degradation. The implementation discussed here is based on BodyChat, a virtual chat system with semi-autonomous avatars.