Characteristics of policy, and experience-driven neural network when applied to level control

The characteristics of the PENN (Policy- and Experience-driven Neural Network) controller were examined by applying it to water-level control. This controller is driven by two learning mechanisms: (1) global learning based on control policy and (2) local learning based on previous experience. The applicability of the PENN controller was tested by simulating water level control of a conical vessel. The neural network used in this study is composed of only four units-three for the input layer and one for the output layer.The effect of continuation of learning, the capability to follow the change of system characteristics, the effect of noise or misjudgment of time lag are discussed.The results found and the fact that this controller has no specific parameters to be tuned in advance showed that it is very adaptive and sufficiently strong against noise.