Design & Validation of ANN based Reinforcement Learning Control Algorithm for Coupled Tank System

This paper presents a framework to apply Reinforcement Learning control algorithm on benchmark nonlinear dynamical systems. This work focuses on a novel Artificial Neural Network (ANN) based dynamic programming approach using Value Iteration to obtain optimal control for continuous-time nonlinear system. In particular, Coupled Tank System has been chosen to represent benchmark nonlinear dynamical system. The proposed Artificial Neural Network-Reinforcement Learning (ANN-RL) algorithm, Naive Reinforcement Learning (Naive-RL) algorithm and traditional PID control schemes are investigated on coupled tank system. The ANN-RL algorithm performs better than the Naive-RL and PID controllers in terms of steady state error, stability, oscillations and overshoot.