A data-driven method for improving a black-box controller

Gisi, Alex

Permalink

https://hdl.handle.net/2142/129310 Copy

Description

Title

A data-driven method for improving a black-box controller

Author(s)

Gisi, Alex

Issue Date

2025-05-05

Director of Research (if dissertation) or Advisor (if thesis)

Norris, William R

Department of Study

Electrical & Computer Eng

Discipline

Electrical & Computer Engr

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Machine learning
• Robotics

Abstract

Optimal control design is an important engineering task. The optimality of a controller is measured by how well closed-loop system trajectories under that controller satisfy a given measure of performance. If a controller is sub-optimal with respect to the performance measure, it would be beneficial to re-design it; however, controller re-design can be expensive. Therefore, it is desirable to alter the closed-loop system performance without touching the baseline controller, that is, by treating it as a black-box. This thesis proposes a method for doing so based on data collected by observing closed-loop trajectories under the baseline controller. The method is based on gain scheduling, or multiplicative modulation of the baseline control signal. A gain scheduling policy describes how and when to apply gains to alter the baseline signal. A gain scheduling policy parameterization and training algorithm for automatically improving the black-box baseline controller is proposed. It is shown that for the proposed policy parameterization, the training can be made more efficient by applying an alternating optimization technique. The resulting gain scheduling policy and training algorithm were applied to three control systems with distinct qualities: a linear time-invariant system, an inverted pendulum, and a skid-steer mobile robot simulation. Additionally, a combined powertrain and kinematic model is developed to implement the mobile robot simulation. To perform realistic evaluations, both linear-quadratic regulator and reinforcement learning-trained controllers are used in the method evaluation. Furthermore, two other learning algorithms beside the one proposed are used to give the results context. It is shown that the proposed method can effectively improve the output of either baseline controller, although certain situations cause worsened performance. The practical implications of these results are examined using examples.

Graduation Semester

2025-05

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/129310

Copyright and License Information

Copyright 2025 Alex Gisi

Owning Collections