FREE ENERGY FORAGING IN AN AFFORDANCE LANDSCAPE

ABSTRACT

Introduction

In this article we construct a simulation of a virtual agent which is equipped with a predictive model of its environment and which operates based on the free energy principle to minimize prediction error. The agent is capable of perceiving a landscape of multiple a_ordances for action in the environment, and selects behaviors towards those a_ordances based on its internal needs and its relation to certain facets of the environment.

Methods

Through the use of a hierarchical model, the agent is endowed with the ability to choose high level behaviors which stabilize its actions, mid level behaviors which respond appropriately to a_ordances, and low level actions which modify action online.

Results

We demonstrate the ability of the agent to engage in foraging behavior based on free energy minimization. The agent is shown to balance multiple conicting needs by responding to appropriate a_ordances in appropriate contexts. The agent is further shown to dynamically adjust its behavior on the y to respond to obstacles in its path. Overall it demonstrates appropriate behavior over multiple timescales.

Conclusions

The agent we have introduced engages in probablistic inference of the hidden states and causes of the external world, and is led by prediction error minimization to either update its model to be more faithful to the dynamics of the world or to execute actions which cause the world to be more in line with its own predictions. Throughout the course of prediction making, the agent is inuenced by the state of the external environment and its internal needs.