Improving Computational Cost of Bayesian Optimization for Controller Tuning with a Multi-stage Tuning Framework

TL;DR

A multi-stage control tuning framework that decomposes control tuning into subtasks, each with a reduced-dimension search space is presented, and it is shown formally that this framework reduces the sample complexity of the control-tuning task.

Abstract

Control auto-tuning for industrial and robotic systems, when framed as an optimization problem, provides an excellent means to tune these systems. However, most optimization methods are computationally costly, and this is problematic for high-dimension control parameter spaces. In this paper, we present a multi-stage control tuning framework that decomposes control tuning into subtasks, each with a reduced-dimension search space. We show formally that this framework reduces the sample complexity of the control-tuning task. We empirically validate this result by applying a Bayesian optimization approach to tuning multiple PID controllers in an unmanned underwater vehicle benchmark system. We demonstrate an 86\% decrease in computational time and 36\% decrease in sample complexity.

Figures (5)

• Figure 1: Control Tuning as a Multistage Optimization Problem
• Figure 2: Control Tuning Multi-Objective Optimization Subtask
• Figure 3: Trajectory Tracking with Independent Tuning
• Figure 4: Trajectory Tracking with Simultaneous Tuning
• Figure 5: XY Plane Response

Theorems & Definitions (2)

• Definition 1: Non-linear Continuous System
• Definition 2: Trajectory Tracking Episode