Adaptive Use of Innovization Principles for a Faster Convergence of Evolutionary Multi-Objective Optimization Algorithms

"Innovization" is a task of learning common principles that exist among some or all of the Pareto-optimal solutions of a multi-objective optimization problem. Except a few earlier studies, most innovization related studies were performed on the final non-dominated solutions found by an EMO algorithm. Since the innovization principles are properties of good and near-optimal solutions, an early identification of them can help improve the evolving population to converge quicker to the Pareto-optimal set. This paper advocates the discovery of innovized principles through machine learning methods during an evolutionary multi-objective optimization run and then using these principles to repair the population adaptively to achieve a faster convergence. Implementing this idea with linear regression as the learning tool and applying it in a test problem with power-law rules existing among Pareto-optimal solutions yields encouraging results. The results show not only an improvement in convergence rate but also in the diversity of non-dominated solutions.