The Past, Present, and Future Challenges of Meta-Black-Box Optimization
As modern optimization tasks grow in number and diversity, automated algorithm design becomes essential. This talk introduces Meta-Black-Box Optimization (MetaBBO), from early ideas to recent advances, and organizes the field into four paradigms: algorithm selection, algorithm configuration, algorithm generation, and solution manipulation. I will focus on our contributions to algorithm generation along three directions: discovering evolutionary update rules in a symbolic mathematical space; assembling optimizers with parameters in a modular space; and leveraging large language models to synthesize executable optimizer code. To strengthen generalization, we study how to build diverse training task sets and how to learn more effective task and state representations. I will also present the MetaBox platform, which aims to improve the development efficiency and support fair evaluation in the field. I will end with a brief look ahead to key opportunities and open questions.
Date and Time
Location
Hosts
Registration
-
Add Event to Calendar
- No. 3333 Liuxian Avenue, Nanshan District
- Shenzhen, Guangdong
- China
- Starts 27 October 2025 04:00 PM UTC
- Ends 07 November 2025 04:00 PM UTC
- No Admission Charge
Media
| 1 | 117.18 KiB |