Communities of Minima in Local Optima Networks of Combinatorial Spaces
In this work we present a new methodology to study the structure of the configuration spaces of hard combinatorial problems. It consists in building the network that has as nodes the locally optimal configurations and as edges the weighted oriented transitions between their basins of attraction. We apply the approach to the detection of communities in the optima networks produced by two different classes of instances of a hard combinatorial optimization problem: the quadratic assignment problem (QAP). We provide evidence indicating that the two problem instance classes give rise to very different configuration spaces. For the so-called real-like class, the networks possess a clear modular structure, while the optima networks belonging to the class of random uniform instances are less well partitionable into clusters. This is convincingly supported by using several statistical tests. Finally, we shortly discuss the consequences of the findings for heuristically searching the corresponding problem spaces.
💡 Research Summary
The paper introduces a novel framework for analysing the structure of combinatorial optimisation landscapes by constructing a Local Optima Network (LON). In a LON, each node corresponds to a locally optimal solution of a problem, and directed weighted edges represent the probability of transitioning from the basin of attraction of one local optimum to another after a single random perturbation followed by local search. This representation compresses the exponentially large search space into a tractable graph that captures the dynamics of stochastic heuristics.
The authors apply the methodology to the Quadratic Assignment Problem (QAP), a classic NP‑hard combinatorial optimisation task. Two families of QAP instances are examined: (i) “real‑like” instances, whose flow and distance matrices are generated with structured correlations that mimic real‑world facility layout problems, and (ii) “uniform random” instances, where matrix entries are drawn independently from a uniform distribution. Both families are generated for comparable problem sizes (typically N = 12–20) to isolate the effect of instance structure.
To build the LON, the authors first generate a large set of random starting configurations and run a simple local search (2‑opt or pairwise exchange) until convergence to a local optimum. Each distinct local optimum becomes a node, and its basin of attraction is identified by recording all configurations that converge to it. Transition probabilities are estimated by repeatedly sampling a random neighbour from a basin, applying the same local search, and noting which basin the resulting solution belongs to. The empirical frequency of each transition defines the edge weight; edges are directed because the probability of moving from basin A to basin B generally differs from the reverse.
Network analysis proceeds with standard graph‑theoretic measures (degree distribution, clustering coefficient, average shortest‑path length) and, crucially, community detection. Three algorithms—Louvain, Infomap, and Walktrap—are employed to uncover modular structure, and the results are compared against null models obtained by randomising edges while preserving degree sequences.
The findings reveal a stark contrast between the two instance classes. Real‑like LONs exhibit high modularity (Q ≈ 0.45–0.55) and a small number of well‑defined communities (typically 5–12). Within each community, edge weights are large, indicating strong intra‑basin transitions, while inter‑community edges are weak, reflecting rare jumps between distinct regions of the landscape. Statistical tests confirm that the observed modularity is significantly greater than that of the randomised null graphs (p < 0.01). By contrast, uniform random LONs display low modularity (Q ≈ 0.15–0.25), diffuse community boundaries, and inconsistent partitions across the three detection methods. Their edge‑weight distribution is near‑uniform, suggesting that the landscape lacks pronounced hierarchical basins.
These structural differences have direct algorithmic implications. In real‑like instances, the presence of tight clusters implies that a heuristic that quickly identifies a community and then intensively explores within it can locate high‑quality solutions efficiently. Strategies that promote occasional long‑range moves (e.g., large mutations, multi‑restart schemes) become valuable for escaping one community and entering another. Conversely, for uniform random instances, the landscape is essentially flat from a modularity perspective; therefore, algorithms must maintain strong global exploration pressure, such as high mutation rates or adaptive diversification mechanisms, to avoid premature convergence.
The authors also analyse the distribution of edge weights to distinguish “core” transitions (high probability) from “peripheral” ones (low probability). Core edges dominate the connectivity of the LON and could be leveraged in a weight‑biased selection rule, whereby a meta‑heuristic preferentially follows high‑probability transitions, potentially accelerating convergence. Peripheral edges, while rare, may serve as escape routes from local traps and could be deliberately sampled to enhance diversification.
Beyond the immediate empirical results, the paper argues that the LON framework is generic and can be extended to other NP‑hard combinatorial problems such as the Traveling Salesman Problem, Graph Partitioning, or Boolean Satisfiability. Future research directions include (1) dynamic LON construction that updates transition estimates on‑the‑fly during a search, (2) the design of LON‑informed meta‑heuristics that adapt their move operators based on community structure, and (3) scalable sampling techniques for approximating LONs in much larger instances where exhaustive basin enumeration is infeasible.
In summary, the study provides a rigorous, network‑theoretic lens for dissecting combinatorial landscapes, demonstrates that instance‑specific structural properties manifest as distinct community patterns in the LON, and shows how these patterns can guide the design of more effective heuristic search strategies.