Effects of training data variation and temporal representation in a QSR-based action prediction system

Understanding of behaviour is a crucial skill for Artificial Intelligence systems expected to interact with external agents - whether other AI systems, or humans, in scenarios involving co-operation, such as domestic robots capable of helping out with household jobs, or disaster relief robots expected to collaborate and lend assistance to others. It is useful for such systems to be able to quickly learn and re-use models and skills in new situations. Our work centres around a behaviourlearning system utilising Qualitative Spatial Relations to lessen the amount of training data required by the system, and to aid generalisation. In this paper, we provide an analysis of the advantages provided to our system by the use of QSRs. We provide a comparison of a variety of machine learning techniques utilising both quantitative and qualitative representations, and show the effects of varying amounts of training data and temporal representations upon the system. The subject of our work is the game of simulated RoboCup Soccer Keepaway. Our results show that employing QSRs provides clear advantages in scenarios where training data is limited, and provides for better generalisation performance in classifiers. In addition, we show that adopting a qualitative representation of time can provide significant performance gains for QSR systems.