Dynamical Model of Action Reinforcement by Gated Working Memory

Abstract

Phasic dopamine response is thought to provide a reward prediction error as a teaching signal for adjustment of an action selection policy. Here we propose a simple neural network dynamical system model based on the dopamine modulated cortical basal ganglia thalamic loops. The system produces a realistic dopamine response and combines it with a gated working memory model, to reinforce the policy when the dopamine signal is above baseline and reverse the policy when below baseline. We illustrate the model with application to the learning of a simple cue action delay reward task.