Generation techniques and a novel on-line adaptation strategy for weight vectors within decomposition-based MOEAs
| Autor: | Antonin Ponsich, Antonio López Jaimes, Alberto Sánchez |
|---|---|
| Rok vydání: | 2019 |
| Předmět: |
010201 computation theory & mathematics
Computer science Lattice (order) 0202 electrical engineering electronic engineering information engineering Evolutionary algorithm Solution set 020201 artificial intelligence & image processing 0102 computer and information sciences 02 engineering and technology 01 natural sciences Multi-objective optimization Algorithm |
| Zdroj: | GECCO (Companion) |
| DOI: | 10.1145/3319619.3322055 |
| Popis: | The success of Multi-Objective Evolutionary Algorithms based on Decomposition (MOEA/D) has generated great interest in the use of a reference set of weight vectors to promote diversity within non-dominated solutions. However, the quality of the solution set obtained heavily depends on the relation between the weight distribution and the Pareto front's shape. This study focuses on a performance comparison of classical techniques for weight vector generation, either based on mixture design or low discrepancy sequences, and a novel approach for updating the weight vectors during the evolutionary process. This approach uses information from the non-dominated individuals to generate weights vectors through a repulsion criterion. Preliminary experiments indicate that this dynamic strategy provides significant benefits when compared to the static Simplex Lattice Design (SLD). |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|