The application of temporal difference learning in optimal diet models

摘要

An experience-based aversive learning model of foraging behaviour in uncertain environments is presented. We use Q-learning as a model-free implementation of Temporal difference learning motivated by growing evidence for neural correlates in natural reinforcement settings. The predator has the choice of including an aposematic prey in its diet or to forage on alternative food sources. We show how the predator's foraging behaviour and energy intake depend on toxicity of the defended prey and the presence of Batesian mimics. We introduce the precondition of exploration of the action space for successful aversion formation and show how it predicts foraging behaviour in the presence of conflicting rewards which is conditionally suboptimal in a fixed environment but allows better adaptation in changing environments.