Tutorial: Machine Learning Assisted Evolutionary Multi-objective Optimization

Machine Learning (ML) provides a powerful synergy with EMO by leveraging the populations generated across different generations as data sets for ML interventions. This allows ML to enhance various phases of the EMO process:

• Problem Structure Learning: ML can identify essential objective sets and perform dimensionality reduction.
• Convergence Enhancement: ML models can guide the creation of new offspring solutions that converge better to the true PF.
• Diversity Enhancement: ML can promote better spread and uniformity of solutions across the PF.
• Pareto Front Analysis: ML can help analyze the Pareto Front, identify natural gaps, and generate new solutions in under-represented regions without requiring full optimization runs.

Foundational studies, such as 'Innovization,' demonstrate how ML can extract innovative design principles from optimal solutions to improve operator design.

Innovized Progress Operators (IPs) are key ML-assisted enhancements. For instance, IP2 focuses on improving convergence by mapping desired transitions along reference vectors in the objective space back to the variable space to guide offspring creation. IP3 focuses on enhancing diversity by identifying under-represented regions or boundaries and generating solutions to fill these gaps, often by leveraging insights from ML models trained on desired solution transitions.

These operators dynamically learn from the evolutionary process, creating targeted interventions to improve the population's characteristics. The Unified Innovized Progress (UIP) operator combines both IP2 and IP3, offering a holistic approach to improve both convergence and diversity simultaneously, adapting its strategy based on the problem's needs and the population's state.