Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Tim N. Nisslbeck, Wouter M. Kouw

TL;DR

It is demonstrated 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.

Abstract

We propose an active inference agent to identify and control a mechanical system with multiple bodies connected by joints. This agent is constructed from multiple scalar autoregressive model-based agents, coupled together by virtue of sharing memories. Each subagent infers parameters through Bayesian filtering and controls 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. It outperforms the uncoupled subagents in terms of surprise and goal alignment.

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

TL;DR

It is demonstrated 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.

Abstract

We propose an active inference agent to identify and control a mechanical system with multiple bodies connected by joints. This agent is constructed from multiple scalar autoregressive model-based agents, coupled together by virtue of sharing memories. Each subagent infers parameters through Bayesian filtering and controls 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. It outperforms the uncoupled subagents in terms of surprise and goal alignment.

Paper Structure

This paper contains 17 sections, 26 equations, 3 figures.

Table of Contents

  1. Introduction
  2. Problem statement
  3. Agent specification
  4. Probabilistic model
  5. Inference
  6. Parameters
  7. Controls
  8. Coupling
  9. Optimization
  10. Experiments
  11. System description
  12. Comparisons
  13. Results
  14. Discussion
  15. Conclusion
  16. ...and 2 more sections

Figures (3)

  • Figure 1: (Left) A double mass-spring-damper system where block $1$ is attached to a stationary frame and block $2$ is attached to the first block. The dynamics of the system are determined by the masses $m_i$ of the blocks, the stiffness of the springs $k_i$, the amount of friction $c_i$ the dampeners provide and gravity $\mathrm{g}$. (Right) A double compound pendulum system consisting of two single compound pendulums joined end-to-end. The dynamics of the system are determined by the masses $m_i$ and lengths $l_i$ of the poles.
  • Figure 2: Comparison of predictions and controls of a set of CARX-EFE agents (top rows) versus a set of uncoupled ARX-EFE agents (bottom rows). Each column represents an agent controlling the first and second mass, respectively.
  • Figure 3: Comparison of model performance of a set of CARX-EFE agents versus a set of uncoupled ARX-EFE agents. Each subplot evaluates a specific aspect of performance: (a) goal alignment and (b) prediction error (surprise). Lower values indicate better performance. The top and bottom row in each subplot show the performance of agents controlling the first and second mass, respectively.