Reinforcement Learning using Gaussian Processes for Discretely Controlled Continuous Systems

MARIANO DE PAULA; ERNESTO C. MARTINEZ

Oro Verde, Entre Ríos

Congreso; RPIC 2011 XIV Reunión de Trabajo en Procesamiento de la Información y Control; 2011

Facultad de Ingeniería - UNER

In many application domains such as production engineering, power electronics and process engineering we find discretely controlled continuous systems (DCCSs) which ones constitute a special subclass of hybrid systems. In this article, we introduce a novel simulation-based approach for DDCS to find a control policy under uncertainty. In the proposed approach we use Reinforcement Learn-ing (RL) with Gaussian Process models (GP) to learn the transitions dynamics descriptive of mode execu-tion and a policy for mode switching. Thus when we have only few samples obtained from the interaction with the real system the dynamic can be drawn by many functions. By Bayesian Inference with GPs is possible to include the uncertainty of the system to make predictions and take control decisions to design an optimal control policy. Throughput maximization and smoothness in a typical PVC production line in the face of significant schedule variability due to resource sharing is used as a case study.