Coupled autoregressive active inference agents for control of multi-joint dynamical systems

Abstract

We propose an active inference agent to learn and control a mechanical system with multiple bodies connected by joints. This agent consists of multiple scalar autoregressive model-based agents that are coupled by virtue of sharing memories. Each agent infers parameters through Bayesian filtering and estimates the most probable control value by minimizing expected free energy over a finite time horizon. We demonstrate that a coupled agent of this kind is able to learn the dynamics of a double mass-spring-damper system and drive it to a desired position through a balance of explorative and exploitative actions. We observe emergence of coordination between the coupled agents.