Stability-constrained reinforcement learning for level control of nonlinear coupled tank system: an experimental study

Stability-constrained reinforcement learning for level control of nonlinear coupled tank system: an experimental study

연구 분야: Infrastructure

학회: Neural Computing and Applications

Neural network (NN) control systems face significant challenges in the theoretical examination of closed-loop stability, despite their success. This paper presents a reinforcement learning (RL) technique with stability constraints for training NN controllers (NNCs) for level control of nonlinear coupled tank systems. The RL adopted in this paper is the deep deterministic policy gradient (DDPG) algorithm. By describing the system in a linear-time-invariant interval system form and utilizing the vertex matrices of the system matrix, a set of Lyapunov-based stability conditions is derived and then employed as training constraints in the DDPG. In each training step, the training constraints are satisfied by utilizing a gradient projection technique. A Lyapunov function, as a result of training, certifies the stability. Simulation and experiment results demonstrate that the constrained NNC works effectively.