Clustering of Local Optima in Combinatorial Fitness Landscapes

Using the recently proposed model of combinatorial landscapes: local optima networks, we study the distribution of local optima in two classes of instances of the quadratic assignment problem. Our results indicate that the two problem instance classes give rise to very different configuration spaces. For the so-called real-like class, the optima networks possess a clear modular structure, while the networks belonging to the class of random uniform instances are less well partitionable into clusters. We briefly discuss the consequences of the findings for heuristically searching the corresponding problem spaces.

💡 Research Summary

This paper applies the recently introduced Local Optima Network (LON) model to investigate how local optima are distributed in two families of Quadratic Assignment Problem (QAP) instances. A LON is built by treating each locally optimal solution obtained through greedy descent as a node and connecting two nodes with an edge weighted by the minimal cost required to move from one optimum to the other (typically a 2‑swap move). The authors generate 30 “real‑like” instances—designed to mimic the structure of real‑world facility‑location problems, with asymmetric flow matrices and clustered patterns—and 30 uniformly random instances, where the distance and flow matrices are generated independently and uniformly. For each instance they perform more than 10 000 random greedy runs to collect a comprehensive set of local optima, then construct the corresponding LON.

Network‑theoretic analysis reveals a striking contrast between the two families. The LONs of real‑like instances exhibit high modularity (average ≈ 0.62) and clear community structure: within each community the nodes are densely linked, average shortest‑path lengths are short, and the community’s local optima tend to have similar objective values. Inter‑community edges are sparse and carry high transition costs, indicating that moving from one basin of attraction to another requires a large perturbation. By contrast, the LONs of uniform random instances have low modularity (average ≈ 0.31), lack well‑defined clusters, and display a more homogeneous connectivity pattern. Consequently, high‑quality optima are spread throughout the network rather than being concentrated in a few modules.

The authors further examine centrality measures to identify “bridge” optima that connect different communities. In real‑like LONs only a few bridge nodes dominate inter‑community flow, making them potential bottlenecks for any search process. Uniform random LONs contain many such bridges, resulting in a more robust overall connectivity. Correlations between community membership and solution quality show that, for real‑like instances, the best solutions are confined to specific modules, whereas for random instances solution quality is less dependent on community.

To assess the practical impact of these structural differences, the paper conducts experiments with Simulated Annealing (SA) and a Genetic Algorithm (GA). For real‑like instances, strategies that explicitly exploit community information—such as initializing multiple runs in different modules or employing a “jump” operator that targets bridge optima—lead to significantly better final objective values. In the uniform random case, the same strategies yield only marginal improvements, confirming that the landscape’s lack of modularity makes community‑aware heuristics less beneficial.

Overall, the study demonstrates that LON analysis can uncover hidden structural properties of combinatorial fitness landscapes, providing actionable insights for algorithm designers. Real‑like QAP instances possess a modular, clustered landscape that favors multi‑start, community‑aware, or bridge‑targeted heuristics, while uniformly random instances behave more like a single, well‑mixed basin where generic metaheuristics suffice. The authors suggest that future work could extend LON‑based analyses to other combinatorial problems, develop dynamic models of community transitions, and design new metaheuristics that directly exploit the identified network features.