Write a PREreview

Reinforcement learning treats reward maximization as a single objective, such that pain avoidance is implicit in pleasure seeking. However, humans appear to have distinct neural systems for processing pain and pleasure. This paper investigates the computational advantages of this separation through grid-world experiments. We demonstrate that modular architectures employing distinct max and min operators for value propagation outperform monolithic models in non-stationary environments. This separation allows agents to simultaneously grow and shrink learned values without interference, enabling both efficient reward collection and punishment avoidance. Additionally, these separate systems can be dynamically arbitrated using a mood-like mechanism for rapid adaptation. Our results suggest that separate pain and pleasure systems may have evolved to enable safe and efficient learning in changing environments.

Nature has placed mankind under the governance of two sovereign masters, pain and pleasure

Jeremy Bentham