Optimization of the Anticipatory Reflexes of a Computational Model of the Cerebellum

Abstract

The cerebellum is involved in avoidance learning tasks, where anticipatory actions are developed to protect against aversive stimuli. In the execution and acquisition of discrete actions we can distinguish errors of omission and commission due to a failure to execute a required defensive Conditioned Response (CR) to avoid an aversive Unconditioned Stimulus (US), and the energy expenditure of triggering an unnecessary CR in the absence of a US respectively. Hence, a motor learning cost function must consider both these components of performance and energy expenditure. Unlike remaining noxious stimuli, unnecessary actions are not directly sensed by the cerebellum. It has been suggested that the Nucleo-Olivary Inhibition (NOI) serves to internally rely information about these needless protective actions. Here we argue that the function of the NOI can be interpreted in broader terms as a signal that is used to learn optimal actions in terms of cost. We work with a computational model of the cerebellum to address: (i) how can the optimum balance between remaining aversive stimuli and preventing effort be found, and (ii) how can the cerebellum use the overall cost information to establish this optimum balance through the adjusting of the gain of the NOI. In this paper we derive the value of the NOI that minimizes the overall cost and propose a learning rule for the cerebellum through which this value is reached. We test this rule in a collision avoidance task performed by a simulated robot.