A "plug and play" computationally efficient approach for control design of large-scale nonlinear systems using cosimulation: A combination of two "ingredients"

Abstract

As mathematical program-ming techniques and com-puter capabilities evolve, control designs are required to address control problems of ever-increasing scale and complexity. A particularly promising methodology toward such a purpose is simulation-based control design (cosimulation). In cosim-ulation, the controller utilizes an op-timizer to minimize or maximize a cost function, related to system per-formance, whose optimization in-volves an iterative process of system simulation and controller redesign. The main applications, which gave a boost to the develop-ment of simulation-based designs, include decision making in manufacturing systems or operational and managerial decision support in other discrete-event pro-cesses (see [1]-[5] and references therein for monographs and survey papers). The advantages of such an approach are many. The controller design does not require any sim-plified or approximated state-space model of the actual system. Moreover, the controller is verified and evaluated using realistic conditions, including physical constraints, delays, and atypical behaviors occurring during the real-life operations of the system. Finally, the control design can be performed in a "plug and play" fashion. The notion of plug and play is that the control design is performed without a tedious and time-consuming analysis of the system properties or of the control design. Instead, the control design is directly driven by input/output data coming from the system, without requiring, for example, any knowledge of a state-space model for the system and its properties. Unfortunately, as simulation-based control design employs optimizers that are called to operate over an elaborate and complex simulation model, their efficien-cy may become problematic, especially when they are applied to large-scale systems (LSS) that involve a large number of states, control inputs, and parameters (see [6]-[10] for some recent large-scale applications). In fact, in most cases, the lack of a mathematical model of the system to be controlled makes the gradient computation of the cost function infeasible. For this reason, gradient descent or similar methods employing Jacobian and Hessian ma-trices [11] are not implementable, and optimization is en-trusted to derivative-free optimization methods, which might show extremely slow convergence in the case of LSS. © 2014 IEEE.