A bio-inspired reinforcement learning model that accounts for fast adaptation after punishment

Abstract

Humans and animals can quickly learn a new strategy when a previously-rewarding strategy is punished. It is difficult to model this with reinforcement learning methods, because they tend to perseverate on previously-learned strategies − a hallmark of impaired response to punishment. Past work has addressed this by augmenting conventional reinforcement learning equations with ad hoc parameters or parallel learning systems. This produces reinforcement learning models that account for reversal learning, but are more abstract, complex, and somewhat detached from neural substrates. Here we use a different approach: we generalize a recently-discovered neuron-level learning rule, on the assumption that it captures a basic principle of learning that may occur at the whole-brain-level. Surprisingly, this gives a new reinforcement learning rule that accounts for adaptation and lose-shift behavior, and uses only the same parameters as conventional reinforcement learning equations. In the new rule, the normal reward prediction errors that drive reinforcement learning are scaled by the likelihood the agent assigns to the action that triggered a reward or punishment. The new rule demonstrates quick adaptation in card sorting and variable Iowa gambling tasks, and also exhibits a human-like paradox-of-choice effect. It will be useful for experimental researchers modeling learning and behavior.