Learning knowledge-level domain dynamics

The ability to learn relational action models from noisy, incomplete observations is essential to support planning and decision-making in real-world environments. While some methods exist to learn models of STRIPS domains in this setting, these approaches do not support learning of actions at the knowledge level. In contrast, planning at the knowledge level has been explored and in some domains can be more successful than planning at the world level. In this paper we therefore present a method to learn knowledge-level action models. We decompose the learning problem into multiple classification problems, generalising previous decompositional approaches by using a graphical deictic representation. We also develop a similarity measure based on deictic reference which generalises previous STRIPS-based approaches to similarity comparisons of world states. Experiments in a real robot domain demonstrate our approach is effective.