Technical Report 2022-006

A Robust Multi-Goal Exploration Aided Tracking Policy

Ruoqi Zhang, Per Mattsson, and Torbjörn Wigren

June 2022

Set-point control aims at finding a policy that can track a set point that varies over time. Such control objectives are central in industry, yet multi-goal Reinforcement Learning methods are typically evaluated on other environments. The paper therefore proposes the use of a combination of feedback based amplitude aided exploration, simulated ensemble model training, together with policy optimization also over integrated errors, to arrive at a trained multi-goal policy that can be directly deployed to real-world nonlinear set-point control systems. The claim is supported by experiments with a real-world nonlinear cascaded tank process and a simulated strongly non-linear pH-control system.

Available as PDF (5.73 MB, no cover)

Download BibTeX entry.