Conference on Cognitive Computational Neuroscience

Rate-distortion theory has been used to model the relationship between the capacity of a resource-limited agent and learning performance. However, most of these models define a single bottleneck on either the representational or policy complexity of the agent, but not both. Here we explicitly model representational capacity and policy capacity separately, and show that they make independent and non-interchangeable impacts on learning performance and efficiency of learned representations. This preliminary work has the potential to provide normative guidance about how to design more efficient RL agents, while also informing better descriptive models of human behavior by capturing different forms of cognitive constraints.