Controller Adaptation via Learning Solutions of Contextual Bayesian Optimization

TL;DR

This paper introduces a framework that uses contextual Bayesian optimization with Gaussian processes to adapt controller parameters for dynamical systems across varying environments, demonstrated on connected vehicle control.

Contribution

It presents a novel approach combining contextual Bayesian optimization and Gaussian processes for controller adaptation in dynamic environments.

Findings

Effective in sim-to-real transfer for vehicle control

Successfully learns optimal controller parameters from simulation

Demonstrates real-time deployment in connected vehicle scenarios

Abstract

In this work, we propose a framework for adapting the controller's parameters based on learning optimal solutions from contextual black-box optimization problems. We consider a class of control design problems for dynamical systems operating in different environments or conditions represented by contextual parameters. The overarching goal is to identify the controller parameters that maximize the controlled system's performance, given different realizations of the contextual parameters.We formulate a contextual Bayesian optimization problem in which the solution is actively learned using Gaussian processes to approximate the controller adaptation strategy. We demonstrate the efficacy of the proposed framework with a sim-to-real example. We learn the optimal weighting strategy of a model predictive control for connected and automated vehicles interacting with human-driven vehicles from simulations and then deploy it in a real-time experiment.