Data-Driven Optimized Tracking Control Heuristic for MIMO Structures: A Balance System Case Study

TL;DR

This paper introduces a data-driven heuristic that optimizes control gains for MIMO systems using neural networks and a nonlinear threshold heuristic, without prior system knowledge, demonstrated on a balance system.

Contribution

It presents a novel control heuristic combining neural networks and threshold accepting for MIMO systems, eliminating the need for prior system dynamics knowledge.

Findings

Effective control gain optimization without system models

Comparable or improved performance over PID-Riccati controller

Validated on a balance system with various optimization criteria

Abstract

A data-driven computational heuristic is proposed to control MIMO systems without prior knowledge of their dynamics. The heuristic is illustrated on a two-input two-output balance system. It integrates a self-adjusting nonlinear threshold accepting heuristic with a neural network to compromise between the desired transient and steady state characteristics of the system while optimizing a dynamic cost function. The heuristic decides on the control gains of multiple interacting PID control loops. The neural network is trained upon optimizing a weighted-derivative like objective cost function. The performance of the developed mechanism is compared with another controller that employs a combined PID-Riccati approach. One of the salient features of the proposed control schemes is that they do not require prior knowledge of the system dynamics. However, they depend on a known region of stability for the control gains to be used as a search space by the optimization algorithm. The control mechanism is validated using different optimization criteria which address different design requirements.