Egocentric representations of the environment have historically been relegated to being used only for simple forms of spatial behaviour such as stimulus-response learning. However, in the many cases that critical aspects of policies are best defined relative to the self, egocentric representations can be advantageous. Furthermore, there is evidence that forms of egocentric representation might exist in the wider hippocampal formation. Nevertheless, egocentric representations have yet to be fully incorporated as a component of modern navigational methods. Here we investigate egocentric successor representations (SRs) and their combination with allocentric representations. We build a reinforcement learning agent that combines an egocentric SR with a conventional allocentric SR to navigate complex 2D environments. We demonstrate that the agent learns generalisable egocentric and allocentric value functions which, even when only additively composed, allow it to learn policies efficiently and to adapt to new environments quickly. Our work shows the benefit for egocentric relational structure to be captured, as well as allocentric. We offer a new perspective on how cognitive maps could usefully be composed from multiple simple maps representing associations between state features defined in different reference frames.
Shani et al. (Fri,) studied this question.