Andrea Stocco - Response Variability in Learning and Decision Making

In this talk I will present the some preliminary results of my recent modeling work that deals with two interrelated problems: (a) The relationship between basal ganglia, learning, and decision-making, and (b) The source of response variability. In particular, I will present a neurocomputational model of the basal ganglia that learns to modulate the variability of its own responses while at the same time learning the correct value of each action. This model naturally balances exploration and exploitation in decision-making, and outperforms a comparable model that shares the same architecture but is based on a traditional neural implementation of a reinforcement learning algorithm. The proposed mechanism is under the dual control of acetylcholine and dopamine, and can be used to account for the peculiar behavioral patterns that are associated with neurological disorders linked to either of these neurotransmitters.

Finally, I will present a prototype of how this mechanism can be scaled up to production system models (like ACT-R), and how it can be used in complex cognitive models.