A Random Forest-Assisted Evolutionary Algorithm for Data-Driven Constrained Multiobjective Combinatorial Optimization...

Area of Science:

• Optimization
• Machine Learning
• Data Science

Background:

• Many real-world optimization challenges lack analytical objective functions, necessitating data-driven solutions.
• Expensive data-driven constrained multiobjective combinatorial optimization problems require specialized methods for evaluation.

Purpose of the Study:

• To develop an effective and efficient algorithm for data-driven constrained multiobjective combinatorial optimization problems.
• To address the challenge of evaluating solutions using only large datasets without analytical functions.

Main Methods:

• Proposed a surrogate-assisted optimization framework utilizing random forests (RFs) and radial basis function networks.
• Incorporated logistic regression models to refine surrogate-assisted fitness evaluations.
• Employed stochastic ranking selection to mitigate the impact of approximated constraint functions.

Main Results:

• Empirical evaluation on multiobjective knapsack and trauma system design problems demonstrated the algorithm's efficacy.
• The variant employing RF models as surrogates proved particularly effective and efficient.
• The proposed methods successfully handled expensive data-driven optimization tasks.

Conclusions:

• The developed surrogate-assisted approach, particularly with RFs, is a viable and efficient strategy for data-driven constrained multiobjective combinatorial optimization.
• This method offers a practical solution for problems where objective functions are not analytically defined.
• The findings have implications for designing complex systems based on extensive data analysis.