Structured Online Learning-based Control of Continuous-time Nonlinear Systems

TL;DR

This paper introduces a structured, online, model-based reinforcement learning method for continuous-time nonlinear systems that efficiently computes optimal control policies using a quadratic value function and matrix differential equations.

Contribution

It presents a novel online learning algorithm leveraging system structure and quadratic value functions to reduce computational complexity in continuous-time nonlinear control.

Findings

Successfully applied to four benchmark nonlinear systems.

Achieved regulation with bounded model prediction error.

Provided a computationally efficient update rule for control parameters.

Abstract

Model-based reinforcement learning techniques accelerate the learning task by employing a transition model to make predictions. In this paper, a model-based learning approach is presented that iteratively computes the optimal value function based on the most recent update of the model. Assuming a structured continuous-time model of the system in terms of a set of bases, we formulate an infinite horizon optimal control problem addressing a given control objective. The structure of the system along with a value function parameterized in the quadratic form provides a flexibility in analytically calculating an update rule for the parameters. Hence, a matrix differential equation of the parameters is obtained, where the solution is used to characterize the optimal feedback control in terms of the bases, at any time step. Moreover, the quadratic form of the value function suggests a compact way of updating the parameters that considerably decreases the computational complexity. Considering the state-dependency of the differential equation, we exploit the obtained framework as an online learning-based algorithm. In the numerical results, the presented algorithm is implemented on four nonlinear benchmark examples, where the regulation problem is successfully solved while an identified model of the system is obtained with a bounded prediction error.