Robotics and the Q-analysis of behaviour

Abstract

A new method is presented for the analysis of complex multi-agent robotic behaviours, based on observing behaviour and abstracting models on which to base control. This involves (i) building an appropriate representation of the scene using the `best' features, (ii) classifying each scene by its features, and (iii) learning the correlation between scene classes and the actions performed in each. A novel aspect is the use of Q-analysis to investigate relational structure between many possible observable features, and congurations in the scene. Q-analysis helps in the selection of appropriate features and is the basis of our combinatorial classification. The methods and algorithms presented are experimentally validated using data from the RoboCup simulation league. We conclude that Q-analysis could be a powerful new approach in robot control.