Neural network with reinforcement learning for adaptive time-optimal control of tank level

摘要

Liquid level control in small standpipes is often a very difficult control problem. An effort is made to control the liquid level in two cascaded tanks having different small time-constants. The control objective for this non-linear system is to minimize the transients in controlled variable and reach new setpoint within optimum time. Simulation studies have been carried out on time-optimal control of a coupled two-tank system using Neural network based controller with Reinforcement learning concept. On-line reinforcement learning has been achieved with adaptation to changes in the plant model. This has been tested by giving sudden change in inflow and valve parameters. User initiated training with adaptation to changes in area of cross-section of the tank, has also been developed. This has been tested by considering tanks of uniform and non-uniform cross-sectional areas. For comparison purposes, simulation studies have also been carried out with (a) PD controller (b) Dynamic programming method based controller and (c) Back propagation algorithm using Neural network.