Abstract:In recent years, graph neural networks (GNNs) have become increasingly popular for solving NP-hard combinatorial optimization (CO) problems, such as maximum cut and maximum independent set. The core idea behind these methods is to represent a CO problem as a graph and then use GNNs to learn the node/graph embedding with combinatorial information. Although these methods have achieved promising results, given a specific CO problem, the design of GNN architectures still requires heavy manual work with domain knowledge. Existing automated GNNs are mostly focused on traditional graph learning problems, which is inapplicable to solving NP-hard CO problems. To this end, we present a new class of \textbf{AUTO}mated \textbf{G}NNs for solving \textbf{NP}-hard problems, namely \textbf{AutoGNP}. We represent CO problems by GNNs and focus on two specific problems, i.e., mixed integer linear programming and quadratic unconstrained binary optimization. The idea of AutoGNP is to use graph neural architecture search algorithms to automatically find the best GNNs for a given NP-hard combinatorial optimization problem. Compared with existing graph neural architecture search algorithms, AutoGNP utilizes two-hop operators in the architecture search space. Moreover, AutoGNP utilizes simulated annealing and a strict early stopping policy to avoid local optimal solutions. Empirical results on benchmark combinatorial problems demonstrate the superiority of our proposed model.
What problem does this paper attempt to address?
### Problems the Paper Aims to Solve
This paper aims to address NP-hard combinatorial optimization (CO) problems, particularly Mixed Integer Linear Programming (MILP) and Quadratic Unconstrained Binary Optimization (QUBO) problems. Specifically, the authors propose an automated Graph Neural Network framework—AutoGNP, for automatically designing optimal Graph Neural Network architectures to improve performance on these NP-hard problems.
### Background and Motivation
Combinatorial optimization problems are crucial in many practical applications, such as transportation and scheduling. However, most combinatorial optimization problems are NP-hard, meaning they are computationally very challenging. Traditional exact solving methods require exponential time, so heuristic methods are often used to find satisfactory suboptimal solutions. In recent years, Graph Neural Networks (GNNs) have been widely applied to solve combinatorial optimization problems and have achieved significant results. Nevertheless, existing GNN architectures still require a lot of manual design and domain knowledge, which limits their performance in tasks.
### Main Contributions
1. **Automated Framework: AutoGNP**
Proposed AutoGNP, an automated Graph Neural Network model specifically for solving combinatorial optimization problems. Additionally, a new differentiable gradient algorithm was designed to identify the optimal Graph Neural Network architecture.
2. **Representation of Combinatorial Optimization Problems**
Focused on two specific combinatorial optimization problems: Mixed Integer Linear Programming and Quadratic Unconstrained Binary Optimization. Demonstrated that these methods are suitable for handling large-scale combinatorial optimization problems.
3. **Experimental Evaluation**
Conducted empirical evaluations of AutoGNP on multiple benchmark datasets, showing that this method outperforms existing GNN architectures in solving NP-hard combinatorial optimization problems.
### Method Overview
- **Graph Representation**
Represented combinatorial optimization problems as graphs, for example, representing MILP instances as weighted bipartite graphs and QUBO problems as ordinary graphs.
- **Graph Neural Architecture Search**
Designed a differentiable graph neural architecture search algorithm, utilizing simulated annealing and strict early stopping strategies to avoid local optima.
- **Experimental Setup**
Conducted experiments on multiple benchmark datasets, including tasks such as set cover, combinatorial auction, facility location, maximum cut, and maximum independent set.
### Experimental Results
- **Maximum Cut Problem**
Experimental results on the Gset dataset show that AutoGNP nearly achieved the best solutions in the field and significantly outperformed the GCN-based baseline method PI-GNN.
- **Other Combinatorial Optimization Problems**
AutoGNP also performed excellently on tasks such as set cover, combinatorial auction, facility location, and maximum independent set, reaching the level of expert solvers.
### Conclusion
This paper proposes an automated Graph Neural Network framework, AutoGNP, for solving NP-hard combinatorial optimization problems. By automatically designing optimal Graph Neural Network architectures, AutoGNP achieved significant performance improvements on multiple benchmark datasets, demonstrating its potential in solving complex combinatorial optimization problems.