An Evolutionary Algorithm with Double-Level Archives for Multiobjective Optimization | JoVE Visualize

Area of Science:

Computational intelligence
Optimization algorithms
Evolutionary computation

Background:

Existing multiobjective evolutionary algorithms (MOEAs) face challenges in solution distribution and population diversity.
Problem decomposition in MOEAs can lead to faster convergence but often relies on a priori decomposition, limiting flexibility.
Current methods struggle with maintaining diversity across sub-problems and ensuring uniform solution distribution along the Pareto front.

Purpose of the Study:

To develop a novel MOEA that combines the strengths of both problem-level and sub-problem-level approaches.
To address limitations in solution distribution and population diversity inherent in existing MOEA strategies.
To enhance the scalability and performance of evolutionary algorithms for multiobjective optimization problems.

Main Methods:

Introduction of a MOEA featuring a double-level archive system: a global archive and multiple sub-archives.
Implementation of a reproduction strategy involving self-reproduction from the global archive and cross-reproduction between global and sub-archives.
Communication and updates between archives facilitated by cross-reproduction to refine solutions.

Main Results:

The proposed algorithm demonstrates fast convergence, similar to problem decomposition methods.
It effectively handles solution distribution along the Pareto front, overcoming limitations of a priori decomposition.
Experimental results on benchmarks and disconnected problems show competitive advantages in distance to Pareto front, solution coverage, and search speed compared to state-of-the-art MOEAs.

Conclusions:

The double-level archive MOEA effectively balances convergence speed with improved solution distribution and diversity.
This novel framework offers a scalable and robust approach to tackling complex multiobjective optimization challenges.
The algorithm presents a significant advancement over existing MOEAs, particularly for problems requiring nuanced Pareto front exploration.