FM4CO contains interesting research papers (1) using Existing Large Language Models for Combinatorial Optimization, and (2) building Domain Foundation Models for Combinatorial Optimization.
Most research utilizes existing LLMs to generate/improve solutions*, algorithms* (hyper-heuristic), or parameters* (hyper-network), yielding impressive results when integrated with problem-specific heuristics or general meta-heuristics. Other studies employ LLMs to investigate the interpretability* of COP solvers, automate problem formulation, or simplify the use of domain-specific tools through text prompts. Given the capabilities of LLMs, this area of research is likely to garner increasing interest. Currently, we mainly focus on the below problems, and may include more variants (e.g., graph-based COPs) as the community grows:
- Traveling Salesman Problem
- Vehicle Routing Problem
- Scheduling Problem
- (Mixed) Integer Linear Programming
Developing a domain foundation model capable of solving a wide range of COPs presents an intriguing and formidable challenge. Recent efforts in this area aim towards this ambitious goal by creating a unified architecture* or representation* applicable across various COPs.
Date | Paper | Link | Problem | Venue | Remark* | Paradigm |
---|---|---|---|---|---|---|
2023.05 | Efficient Training of Multi-task Combinatorial Neural Solver with Multi-armed Bandits | TSP,VRP, OP,KP |
ArXiv | Architecture | ||
2024.02 | Multi-Task Learning for Routing Problem with Cross-Problem Zero-Shot Generalization | VRP |
KDD 2024 | Architecture | Compositional Zero-Shot | |
2024.03 | Towards a Generic Representation of Combinatorial Problems for Learning-Based Approaches | SAT,TSP, COL,KP |
ArXiv | Representation | ||
2024.04 | Cross-Problem Learning for Solving Vehicle Routing Problems | TSP ,OP , PCTSP |
IJCAI 2024 | Architecture | Efficient Fine-Tuning | |
2024.05 | MVMoE: Multi-Task Vehicle Routing Solver with Mixture-of-Experts | VRP |
ICML 2024 | Architecture |