Evolutionary Multi and Many-Objective Optimization: Enhancements Using Machine Learning

Evolutionary Algorithms (EAs) are known to efficiently approximate the Pareto-front (PF) in a single simulation run for problems with conflicting objectives. Conventionally, problems with two- or three-objectives are referred to as multi-objective problems, while those with four or more objectives are referred to as many-objective problems. EAs capable of handling both these problem classes are referred to as Evolutionary Multi- and Many-objective Optimization Algorithms (EMâOAs). Despite their success, researchers continually challenge EMâOAs with increasingly complex real-world problem characteristics, where ensuring convergence to and diversity across the Pareto-optimal front (PF) becomes difficult. Most efforts to improve EMâOAs’ efficiency have focused on improving the inherent bio-inspired operators. However, treatment of EMâOAs’ generational populations as data sets amenable to Machine Learning (ML) towards alleviating converge and diversity challenges is an emergent trend. As proponents of the latter, the authors’ research has demonstrated that ML techniques can utilize the generational non-dominated solutions to (i) discern the underlying problem structure, potentially facilitating dimensionality reduction, (ii) learn efficient search trajectories to expedite convergence towards Pareto-optimal (PO) solutions, (iii) learn mechanisms to improve diversity among PO solutions, and (iv) generate new non-dominated solutions from the final PO solutions to help find the preferred solution by the decision-maker. This chapter, while introducing a holistic perspective, dominantly focuses on presenting how ML techniques can enhance the search efficiency of EMâOAs in terms of convergence and diversity. The authors envision that the synergy between EMâO and ML could be utilized further to enrich each of these domains, for tacking complex real-world problems.