FIMP-HGA: A Novel Approach to Addressing the Partitioning Min-Max Weighted Matching Problem

Yuxuan Wang,Jiongzhi Zheng,Jinyao Xie,Kun He

2024-05-06

Abstract:The Partitioning Min-Max Weighted Matching (PMMWM) problem, being a practical NP-hard problem, integrates the task of partitioning the vertices of a bipartite graph into disjoint sets of limited size with the classical Maximum-Weight Perfect Matching (MPWM) problem. Initially introduced in 2015, the state-of-the-art method for addressing PMMWM is the MP$_{\text{LS}}$. In this paper, we present a novel approach, the Fast Iterative Match-Partition Hybrid Genetic Algorithm (FIMP-HGA), for addressing PMMWM. Similar to MP$_{\text{LS}}$, FIMP-HGA divides the solving into match and partition stages, iteratively refining the solution. In the match stage, we propose the KM-M algorithm, which reduces matching complexity through incremental adjustments, significantly enhancing runtime efficiency. For the partition stage, we introduce a Hybrid Genetic Algorithm (HGA) incorporating an elite strategy and design a Greedy Partition Crossover (GPX) operator alongside a Multilevel Local Search (MLS) to optimize individuals in the population. Population initialization employs various methods, including the multi-way Karmarkar-Karp (KK) algorithm, ensuring both quality and diversity. At each iteration, the bipartite graph is adjusted based on the current solution, aiming for continuous improvement. To conduct comprehensive experiments, we develop a new instance generation method compatible with existing approaches, resulting in four benchmark groups. Extensive experiments evaluate various algorithm modules, accurately assessing each module's impact on improvement. Evaluation results on our benchmarks demonstrate that the proposed FIMP-HGA significantly enhances solution quality compared to MP$_{\text{LS}}$, meanwhile reducing runtime by 3 to 20 times.

Neural and Evolutionary Computing

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the Partitioning Min - Max Weighted Matching (PMMWM) problem, which is an NP - hard problem. Specifically:

- **Problem Background**: The PMMWM problem involves partitioning the vertex set \(U\) into several disjoint subsets (each subset contains at most \(\bar{u}\) vertices) in a weighted bipartite graph \(G(U, V, E)\), and simultaneously solving the maximum matching so as to minimize the weight of the heaviest subset. Here, the weight refers to the sum of the weights of the edges matched into this subset.

- **Limitations of Existing Methods**: The current state - of - the - art method is the MP LS algorithm. It uses the KM algorithm in the matching phase, the local search method in the partitioning phase, and iteratively optimizes the quality of the solution. However, MP LS still has room for improvement in the search efficiency in the matching phase and the search ability in the partitioning phase.

- **Contributions of the Paper**: In order to solve the PMMWM problem and overcome the limitations of MP LS, this paper proposes a new method - Fast Iterative Match - Partition Hybrid Genetic Algorithm (FIMP - HGA). FIMP - HGA adopts a fast iterative framework, combines the exact matching algorithm and the hybrid genetic algorithm, and introduces the elitist strategy to improve the quality and efficiency of the solution.

### Specific Problem Description

The PMMWM problem can be formalized as the following optimization problem:

\[
\min_{z} \max_{k \in \{1, \ldots, m\}} \left\{ \sum_{u \in U} \sum_{v \in V} w(e_{uv}) z_{uvk} \right\}
\]

where:
- \(z_{uvk}\) is a binary variable, indicating that if the vertex \(u \in U_k\) and the edge \(e_{uv} \in \Pi\), then \(z_{uvk} = 1\), otherwise \(z_{uvk} = 0\).
- The constraints include:
- Each vertex \(u \in U\) must belong to exactly one subset \(U_k\).
- The size of the subset cannot exceed \(\bar{u}\).
- The matching must be a maximum matching.

### Solution

In order to effectively solve this problem, FIMP - HGA adopts the following strategies:

1. **Matching Phase**: Use the KM - M algorithm, which is an improved version based on the KM algorithm and can reduce redundant calculations in each iteration, thereby reducing the complexity from \(O(n^3)\) to \(O(n^2)\).

2. **Partitioning Phase**: Introduce the Hybrid Genetic Algorithm (HGA), which combines the genetic algorithm and the local search method to explore a broader and deeper solution space. HGA also includes the elitist strategy to ensure that the optimal individuals in each generation can survive.

3. **Graph Modification Phase**: At the end of each iteration, adjust the matching solution of the bipartite graph by disabling some edges and re - enabling these edges in subsequent iterations, which helps the algorithm explore a broader solution space.

Through these innovations, FIMP - HGA not only improves the quality of the solution but also significantly reduces the running time, achieving a 3 - to - 20 - fold speedup compared to MP LS.

### Summary

The main objective of this paper is to solve the PMMWM problem by proposing FIMP - HGA, especially making significant progress in the optimization of the matching and partitioning phases.