Zero-Shot DDPG Controller Design for Liquid Level Control of a Benchmark Quadruple-Tank Process

MacHacuay J.; Ipanaqué W. · 10th 2024 International Conference on Control, Decision and Information Technologies, CoDIT 2024 (2024) · DOI: 10.1109/codit62066.2024.10708234

Resumen

Deep Reinforcement Learning (DRL) methods are increasing the attention of the research community to control complex multivariable nonlinear systems with interactions. This research work will design a Deep Deterministic Policy Gradient (DDPG) controller for a benchmark Quadruple-Tank Process. In addition, a primary concern regarding DRL-based methods is how well they can generalize their experiences to perform well on never-seen-before tasks (i.e., zero-shot generalization) in the same environment. The proposed DDPG controller design addresses this issue simply and efficiently. The results show empirical evidence that the DDPG controller has smooth reference tracking with low overshoots and no oscillations, without requiring any model and just learning from the actual data of the system. Moreover, the described behavior was exhibited in a multiple set point trajectory meanwhile the controller was trained in a single set point trajectory, thus demonstrating the zero-shot design.

Instituciones: Universidad de Piura