A computational model of sequential movement learning with a signal mimicking dopaminergic neuron activities

We present a computational model of approach learning in a simulated maze environment. Our maze environment and training method mimics those used in the experimental literature. We show that our model learns the correct sequence of six decisions that lead to the location of positive reinforcement and in a manner consistent with experimental observations. Our model exhibits many properties that are characteristic of animal learning in maze environments including delay conditioning, secondary conditioning, and backward chaining. Finally, we map our model to the basal ganglia and show that a signal in our model that is responsible for learning has the same temporal properties as dopamine, the neurotransmitter believed to play an important part in learning decision sequences.