Parameterless Gene-Pool Optimal Mixing Evolutionary Algorithms

Area of Science:

• Artificial Intelligence
• Computational Optimization
• Evolutionary Computation

Background:

• Evolutionary algorithms (EAs) are powerful tools for optimization, but their performance often hinges on effectively detecting and exploiting variable dependencies (linkage).
• The gene-pool optimal mixing evolutionary algorithm (GOMEA) is designed to address linkage learning, but further enhancements are needed for broader applicability and improved performance.
• Existing linkage-aware EAs, such as DSMGA-II, offer competitive solutions but may not fully capture complex dependency structures.

Purpose of the Study:

• To present an enhanced version of GOMEA, optimized through large-scale design-space search.
• To introduce a novel variant, CGOMEA, which improves linkage exploitation by incorporating conditional dependency filtering.
• To evaluate the performance of the enhanced GOMEA and CGOMEA against DSMGA-II on challenging black-box optimization problems and investigate parameterless operation via automatic population management.

Main Methods:

• Extensive empirical evaluation of GOMEA and CGOMEA on nine benchmark black-box problems requiring linkage discovery.
• Comparison with a leading linkage-aware EA, DSMGA-II, to assess relative performance.
• Investigation of automatic population management schemes to enhance the usability and robustness of GOMEA and CGOMEA, aiming for parameterless operation.

Main Results:

• The optimized GOMEA and the novel CGOMEA demonstrate significant performance improvements over the original GOMEA and DSMGA-II across most tested problems.
• CGOMEA's linkage-based variation, enhanced by conditional dependency filtering, proves particularly effective.
• Automatic population management schemes enable GOMEA and CGOMEA to achieve competitive results with reduced parameter tuning, enhancing their practical applicability.

Conclusions:

• The enhanced GOMEA and CGOMEA represent a new state-of-the-art in linkage-aware evolutionary computation for complex optimization problems.
• Exploiting conditional dependencies offers a promising avenue for further improving linkage learning in EAs.
• The development of parameterless EAs through automatic population management increases their accessibility and reliability for practitioners.