Learning Subjective Representations for Planning

Planning involves using a model of an agent's actions to find a sequence of decisions which achieve a desired goal. It is usually assumed that the models are given, and such models often require expert knowledge of the domain. This paper explores subjective representations for planning that are learned directly from agent observations and actions (requiring no initial domain knowledge). A non-linear embedding technique called Action Respecting Embedding is used to construct such a representation. It is then shown how to extract the effects of the agent's actions as operators in this learned representation. Finally, the learned representation and operators are combined with search to find sequences of actions that achieve given goals. The efficacy of this technique is demonstrated in a challenging robot-vision-inspired image domain.