Digital Chemical Engineering (Dec 2024)
Parallelizing process model integration for model predictive control through oracle design and analysis for a Grover’s algorithm-inspired optimization strategy
Abstract
In model predictive control (MPC), a process dynamic model is utilized to make predictions of the value of the objective function and constraints throughout a prediction horizon. In one method of solving this problem, the time required to find the optimal values of the decision variables depends on the time required to perform the arithmetic operations involved in computing the model predictions. Methods for attempting to reduce the computation time of an MPC could then include developing approximate (reduced-order or data-driven) models for a system that take less time to solve, or to parallelize the computations using, for example, multiple cores or CPU’s. However, an observation in all of these cases is that the values of the process states across the prediction horizon are not the values returned by the optimization problem; the manipulated input trajectory is the desired decision variable. An optimization strategy that cannot explicitly return the process states but can generate some representation of them that then leads to computation of the desired process input would thus be suitable for MPC. Quantum computers achieve their parallelism through creating values that cannot all be returned. They can then operate on this set of values to return a number that is meaningful with respect to that set of values that could not all be returned. Motivated by this, we wish to investigate an idea for utilizing quantum parallelism in developing a representation of an objective function that depends on the solution of a process dynamic model, but then only returning the control actions that minimize the objective function value dependent on those solutions. To do this, several steps are necessary. The first is to locate a quantum algorithm which has the desired characteristics for achieving the goals. In this work, we perform these steps using an amplitude amplification strategy based on Grover’s algorithm on a quantum computer. The second is to analyze the algorithm with respect to its ability to translate across problems in the MPC domain, with respect to both its ability to handle nonlinear systems and to handle a variety of different structures of the set of all possible objective function values given the allowable values of the decision variables. We thus evaluate the benefits and limitations of the algorithm from this perspective. We do not wish to imply that this algorithm is more computationally-tractable for use with MPC than classical optimization techniques traditionally applied in an MPC context. Rather, we wish to understand the manner in which such an algorithm would be designed and when it is appropriate for MPC problems (in the sense of returning the correct answers to the optimization problem), as a step toward better understanding the interactions of the quantum properties of quantum algorithms with control goals. We also see this as forming an important first step in algorithm design/analysis, which can then translate to future works in comparing this technique computationally with classical “competitor” algorithms to guide further work in searching for relevant quantum algorithms for control applications.
