Addressing real-time control problems in complex environments using dynamic multi-objective evolutionary approaches

TL;DR: A framework for on-line optimisation of dynamic problems that is capable of representing problems in a quantitative way, identifying optimal solutions using multi-objective evolutionary algorithms, and automatically selecting an optimal solution among alternatives is developed.

Abstract: The demand for increased automation of industrial processes generates control problems that are dynamic, multi-objective and noisy at the same time. The primary hypothesis underlying this research is that dynamic evolutionary methods could be used to address dynamic control problems where con icting control criteria are necessary. The aim of this research is to develop a framework for on-line optimisation of dynamic problems that is capable of a) representing problems in a quantitative way, b) identifying optimal solutions using multi-objective evolutionary algorithms, and c) automatically selecting an optimal solution among alternatives. A literature review identi es key problems in the area of dynamic multi-objective optimisation, discusses the on-line decision making aspect, analyses existing Multi- Objective Evolutionary Algorithms (MOEA) applications and identi es research gap. Dynamic evolutionary multi-objective search and on-line a posteriori decision maker are integrated into an evolutionary multi-objective controller that uses an internal process model to evaluate the tness of solutions. Using a benchmark multi-objective optimisation problem, the MOEA ability to track the moving optima is examined with di erent parameter values, namely, length of pre-execution, frequency of change, length of prediction interval and static mutation rate. A dynamic MOEA with restricted elitism is suggested for noisy environments.To address the on-line decision making aspect of the dynamic multi-objective optimisation, a novel method for constructing game trees for real-valued multiobjective problems is presented. A novel decision making algorithm based on game trees is proposed along with a baseline random decision maker. The proposed evolutionary multi-objective controller is systematically analysed using an inverted pendulum problem and its performance is compared to Proportional{ Integral{Derivative (PID) and nonlinear Model Predictive Control (MPC) approaches. Finally, the proposed control approach is integrated into a multi-agent framework for coordinated control of multiple entities and validated using a case study of a tra c scheduling problem.