Planning as Inference in a Hierarchical Predictive Memory

Abstract

In the predictive brain hypotheses, the functional mechanism of the brain is suggested to infer the cause of current states within the predictions by brains. Recently, there have been several approaches to explain the action generation within the predictive brain hypotheses: the brain predicts the animal’s own action, which the animal realizes to fulfill the prediction. In this study, we suggest a predictive brain models to produce the goal directed behaviors. We introduced the Planning as Inference (PAI) framework to a hierarchical predictive memory model. PAI is a computational framework for goal-directed behavior generation. PAI explains the decision of an action for a state in the probabilistic distribution. The distribution is inferred from the evidences of current state and the perspective evidence of goal achievement. We used a hierarchical predictive memory system to predict the agent’s self-action states. Following to the PAI, the predictions were inferred from the evidence of the ongoing state and the evidence from assumption of the goal achievement. The agents realizes the predicted actions to minimize prediction errors. We implemented our method in embodied robotics system and our model could generate structured spontaneous behavior and goal directed behaviors. Our result opens understanding for the goal-directed behavior in predictive brain hypotheses.