Learning interaction structure using a hierarchy of dynamical systems

Abstract

The IAM (Interaction Adaptation Manager) algorithm was recently proposed to learn the optimal parameters of a hierarchical dynamical system incrementally through interacting with other agents given that the structure of the system is known (the number of processes in each layer and their interconnections) and that the agent knows how to interact in all roles except the one it is learning (e.g. an agent learning to listen should know how to speak). This paper presents an algorithm for learning the structure of a hierarchical dynamical system representing the interaction protocol at various timescales and using multiple modalities relaxing these two constraint. The proposed system was tested in a simulation environment in which rich human-like agents are interacting and showed accurate recognition of the interaction structure using few training examples. The learned structure showed acceptable performance that allowed subsequent application of the adaptation algorithm to converge to a good solution using as few as 15 interactions. © 2009 Springer-Verlag Berlin Heidelberg.