Adaptive critic design applied to constraint optimization

A novel technique for handling constraints in adaptive critic design (ACD) is proposed. The technique is applied to the optimization of a simplified alkylation process. Historically, the constraints are embedded within the objective function as a penalty function. This modified unconstrained objective function is then minimized using the ACD architecture. The critic in the ACD architecture predicts the value of the unconstrained objective function, and trains the action network to minimize it. We suggest an enhanced architecture of ACD, which involves two critics instead of one. The first critic (cost critic) outputs the long-term estimate of the cost function and the second critic (constraint critic) calculates the constraint violation. The interactions of these two critics are studied on the training of the action network and compared with the traditional penalty function based approach. The outcomes based on two critics appear to be better as compared to those obtained from a single critic estimating a single objective with constraints embedded as a penalty function.