LLM-Evolved Heuristics Enable Real-Time Adaptive Scheduling
DSevolve, an innovative framework, leverages large language models to develop adaptive scheduling rules tailored for dynamic flexible job shops. It distinguishes between the offline creation of a rule library and the online selection of rules based on current states. By employing an LLM-guided quality-diversity search, this system integrates multi-persona seeding, a MAP-Elites behavioral archive, and behavior-driven variations to generate complementary rules. This methodology effectively tackles issues such as order arrivals, machine failures, and variations in processing times, all while ensuring swift responses for online rescheduling.
Key facts
- DSevolve is a dynamic self-evolutionary framework for scheduling.
- It uses LLM-guided quality-diversity search.
- Offline rule-library construction is separated from online selection.
- The system handles order arrivals, machine breakdowns, and processing-time deviations.
- Multi-persona seeding and MAP-Elites archive are employed.
- Behavior-guided variation evolves complementary rules.
- The framework aims for real-time adaptive scheduling.
- It is designed for dynamic flexible job shops.
Entities
Institutions
- arXiv