Curbing Negative Influences Online for Seamless Transfer Evolutionary Optimization.

Area of Science:

• Computational Intelligence
• Optimization Theory
• Machine Learning

Background:

• Human ability to leverage past experiences for new tasks.
• Limitations in current black-box optimization due to lack of problem-specific data.
• Need for enhanced computational solvers that mimic human learning.

Purpose of the Study:

• To develop a computational framework for online learning and exploitation of similarities across optimization problems.
• To achieve an algorithmic realization of the transfer optimization paradigm.
• To address the challenge of limited prior data in black-box optimization settings.

Main Methods:

• Introduction of a novel evolutionary computation framework.
• Enabling online learning and exploitation of cross-problem similarities.
• Accounting for latent similarities revealed during evolutionary search.
• Theoretical analysis of the proposed framework's influence on performance.

Main Results:

• Theoretical analysis substantiates positive influences on optimization performance.
• Demonstration of practical efficacy via an adaptive transfer evolutionary algorithm instantiation.
• Validation across diverse numerical examples: discrete, continuous, single- and multi-objective optimization.

Conclusions:

• The proposed framework effectively enables online learning and exploitation of problem similarities.
• Adaptive transfer evolutionary algorithms show practical efficacy in enhancing optimization.
• The approach offers significant performance enhancements for global black-box optimizers.