Abstract
Many real-world problems involve multi-objective optimization. Although traditional multi-objective evolutionary algorithms (MOEAs) have achieved considerable success in solving certain types of multi-objective problems (MOPs), they often encounter difficulties when dealing with problems exhibiting multiple complex features, and struggle to find high-quality non-dominated solutions. To address this issue, this paper proposes a reinforcement learning-based hybrid evolutionary multi-objective optimization framework (RL-MOEA), designed to handle MOPs with mixed features. RL-MOEA integrates reinforcement learning with MOEAs to improve adaptability during optimization. This framework models the evolutionary process as a Markov Decision Process (MDP), employing two state representations to capture population distribution and using hypervolume (HV) as a reward signal. Through reinforcement learning, intelligent agents are trained to dynamically select the most suitable optimal method at different evolutionary stages. Computational experiments indicate that RL-MOEA outperforms several state-of-the-art MOEAs on various benchmark problems, achieving optimal inverted generational distance (IGD) values on MOPs with mixed features. The proposed method also exhibits better convergence and diversity under comparable computational budgets. These results demonstrate the capability of RL-MOEA to efficiently solve a wide range of complex MOPs.
| Original language | English |
|---|---|
| Article number | 114318 |
| Journal | Applied Soft Computing Journal |
| Volume | 187 |
| DOIs | |
| State | Published - Feb 2026 |
Keywords
- Deep Q-Network
- Evolutionary algorithms
- Multi-objective optimization
- Reinforcement learning
Fingerprint
Dive into the research topics of 'A hybrid evolutionary framework assisted by reinforcement learning for mixed multi-objective optimization features'. Together they form a unique fingerprint.Cite this
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver