Variable Neighborhood Search (VNS)

• VNS is a metaheuristic for global and combinatorial optimization that systematically changes neighborhood structures to overcome local optima.
• It operates with a structured cycle of shaking, local search, and neighborhood change to balance intensification and diversification.
• VNS has shown strong performance across diverse domains such as scheduling, routing, and bioinformatics, often outperforming alternative heuristics.

Variable Neighborhood Search (VNS) is a general-purpose metaheuristic designed for combinatorial and global optimization, distinguished by its explicit strategy of exploiting multiple neighborhood structures to systematically escape local optima. VNS was introduced by Mladenović and Hansen (1997) and has become foundational across a range of application domains, including bioinformatics, scheduling, routing, clustering, and graph partitioning. Its central premise is that different neighborhood structures typically yield distinct local minima, and that purposeful exploration—by changing neighborhoods during search—enables effective intensification and diversification.

1. Framework and Core Algorithmic Structure

VNS operates through a cycle of three phases: shaking, local search, and neighborhood change. Let $x$ denote the current solution and let $\mathcal{N} = \{N_1, N_2, ..., N_k\}$ be a collection of neighborhood structures.

1. Shaking Phase: Randomly perturb $x$ within the $k$-th neighborhood $N_k$ to obtain $x'$, aiming to move out of the current basin of attraction.
2. Local Search Phase: Apply an intensification method (greedily or otherwise) restricted to $N_k$ starting from $x'$, resulting in a new solution $x''$ at a local optimum of $N_k$.
3. Neighborhood Change:
   • If $\mathcal{N} = \{N_1, N_2, ..., N_k\}$0 improves the objective, accept $\mathcal{N} = \{N_1, N_2, ..., N_k\}$1 and reset $\mathcal{N} = \{N_1, N_2, ..., N_k\}$2 (returning to the most focused neighborhood).
   • Otherwise, increment $\mathcal{N} = \{N_1, N_2, ..., N_k\}$3, broadening the search (exploring larger/different neighborhoods).

This cycle continues until a stopping condition (maximum iterations, time limit, or convergence) is met. Additional mechanisms may include adaptive or probabilistic acceptance of new solutions, multi-start procedures, or randomized neighborhood orderings.

2. Neighborhood Structures and Shaking Mechanisms

Neighborhoods define the local moves available from any solution. In VNS, neighborhoods are purposely diverse in type and "radius," with larger $\mathcal{N} = \{N_1, N_2, ..., N_k\}$4 generally indicating increasingly disruptive moves (greater diversification).

• Partitioning and Clustering Problems: Neighborhood $\mathcal{N} = \{N_1, N_2, ..., N_k\}$5 may represent all solutions differing in the cluster assignment of exactly $\mathcal{N} = \{N_1, N_2, ..., N_k\}$6 elements, with $\mathcal{N} = \{N_1, N_2, ..., N_k\}$7 controlling the amplitude of perturbation. For example, in $\mathcal{N} = \{N_1, N_2, ..., N_k\}$8-plex partitioning, $\mathcal{N} = \{N_1, N_2, ..., N_k\}$9 is constructed by reassigning $x$0 vertices to existing or new components (Grbić et al., 2018).
• Label-Selection/Subset Problems: Neighborhoods are often defined using the Hamming distance; $x$1 is the set of label sets differing in exactly $x$2 labels relative to $x$3 (Consoli et al., 2015).
• Scheduling/Sequencing: Neighborhoods correspond to permutation-based operators (e.g., 2-opt, block shifts, exchange, inversion). Multiple neighborhoods can be assembled depending on the move operator; for instance, seven classes in the single-machine total weighted tardiness problem (Geiger et al., 2011).
• Vehicle Routing and Assignment: Neighborhoods may act on route structures (e.g., multi-customer exchanges, 2-opt*, segment relocation, inter-route swaps) (Detti et al., 2016, 2505.23098). The size and type of neighborhood ($x$4) balance intensification and diversification.

The shaking mechanism is typically a randomized application of one or more moves from $x$5. For challenging combinatorial spaces, the amplitude $x$6 or $x$7 is dynamically tuned, either statically, in proportion to the solution size, or via reactive/adaptive strategies (Consoli et al., 2015, Consoli et al., 2015).

3. Local Search and Intensification Strategies

After shaking, VNS incorporates a dedicated local search subroutine to reach a local optimum with respect to the current neighborhood. The literature reports several variants:

• First-Improvement/1-Swap: Scan candidate moves; accept the first improving one, allowing rapid escape from shallow local minima (Grbić et al., 2018).
• Best-Improvement/Descent: Evaluate all neighbors, select the best, and repeat until no further improvement is found. Efficient partial updates (e.g., for edge-weighted problems) reduce computational complexity.
• Variable Neighborhood Descent (VND): Employ ordered or adaptive sequences of multiple neighborhoods within the local search phase itself (Matic et al., 2015, Geiger et al., 2011). Adaptive VND may use lookahead to select the most promising neighborhood based on sampled improvements.
• Exact or Heuristic Intensifiers: For clustering, a hierarchical agglomerative method (Ward's-GC) or Lloyd's $x$8-means can operate as the intensifier (Martins, 2017).
• Domain-Specific Adjustments: In VRP-type problems, constraint-specific repairs (e.g., to obtain feasibility with respect to capacity or time windows) may be integrated after local search (Detti et al., 2016).

Acceptance criteria usually combine improvement in the primary objective, feasibility, and possibly further lexicographic or stochastic acceptance rules (e.g., probabilistic acceptance of equal-quality solutions to encourage exploration).

4. Parameter Control, Adaptation, and Algorithmic Variations

Parameterization in VNS affects the breadth and efficiency of exploration. Standard approaches include:

• Neighborhood Indexing: Range of $x$9 or $k$0: statically fixed, proportional to problem size or solution diversity, or adaptively updated with search progress (Consoli et al., 2015, Consoli et al., 2015).
• Neighborhood Selection Rules: Sequential, random, or adaptive lookahead. Empirical studies show adaptive selection accelerates convergence when evaluation budgets are tight, while fixed order may provide superior search-depth in moderate-to-large runs (Geiger et al., 2011).
• Hybridization and Learning-Enhanced VNS: Modern extensions employ multi-armed bandit policies for neighborhood selection or deep reinforcement learning (e.g., via transformer policies) to guide both the shaking phase and operator choice, yielding improved generalization and efficiency (2505.23098).

Several application-driven VNS variants supplement the core scheme with problem-specific enhancements—such as complementary solution-space exploration, simulated annealing–style probabilistic local search, or dynamic amplitude control—all contributing to increased performance without intensive user-tuning (Consoli et al., 2015).

5. Applications in Combinatorial Optimization

VNS has demonstrated versatility and effectiveness across diverse problem classes:

Domain	Application Examples	VNS Structure Features
Biological Network Partitioning	Maximum edge-weighted $k$1-plex partitioning	$k$2-swap assignments, custom feasibility/weight objective, 1-swap local search (Grbić et al., 2018)
Bin Packing	Bin Packing with Compatible Categories	Multiple removal/repacking neighborhoods, adaptive selection, two-stage local search (Santos et al., 2019)
Scheduling/Sequencing	Bandwidth coloring, multi-objective flowshop scheduling	Shake on increasing cardinality, VND or randomized neighborhood order (Matic et al., 2015, 0809.0271)
Graph Labeling and Cover	MLST, k-LSF, feature-selection, goal clustering	Hamming neighborhoods, shaking amplitude proportional to $k$3, hierarchical/agglomerative descent
Routing and Dial-a-Ride	Multi-depot DARP, vehicle routing with multiple time windows	Swap, chain, relocate, eliminate, local route search, feasibility repair (Detti et al., 2016, 2505.23098)
Assignment and Generalized TSP	Multidimensional assignment problems, orienteering with sets	VLSN, grid-based restart, multi-start VNS, shake via segment-exchange (Kammerdiner et al., 2021, Kant et al., 2024)
Clustering (Goal-based)	Partitioning with $k$4 constraint, cluster minimization	Point-to-singleton shaking, Ward's-GC/kmeans-GC local search (Martins, 2017)
Drone/TSP Hybrid Logistical Models	Flying Sidekick TSP, truck+drone synchronization	Seven neighborhoods, best-improvement randomized VND, sequence of truck/drone moves (Freitas et al., 2018)

Problem-specific objective functions, feasibility criteria, and neighborhood encodings are essential to the effectiveness and performance of VNS in each of these settings.

6. Computational Characteristics and Empirical Performance

Across multiple domains, VNS exhibits several robust empirical properties:

• Solution Quality: Consistently matches or outperforms alternative heuristics such as genetic algorithms, GRASP, and greedy methods on benchmark sets; for many problems, it reproduces all known optima or improves existing bounds (Grbić et al., 2018, Kant et al., 2024).
• Scalability: Handles large instances (e.g., $k$5 for $k$6-plex partitioning, $k$7 for orienteering and VRP) with practical computation times—often minutes to hours—using local or incremental updates (Grbić et al., 2018, Kant et al., 2024).
• Diversity–Intensification Balance: Dynamic or state-dependent control of shaking amplitude ($k$8, $k$9) enables traversal of large, multimodal search spaces without premature convergence.
• Algorithmic Efficiency: Efficient local objective recalculation, partial updates, and targeted neighborhood scans accelerate the search, especially in large-scale or sparse instances (Grbić et al., 2018, Santos et al., 2019).
• Parameter Tuning: Many advanced VNS implementations are parameter-free or require only mild instance-dependent calibration; some employ full automation of parameter selection via online adaptive rules (Consoli et al., 2015, 2505.23098).

Empirical studies highlight that no single neighborhood operator dominates across all instances, and that combining several strong neighborhoods as in MOVNS (randomized VNS) or using adaptive selection greatly enhances final solution diversity and quality (0809.0271, Geiger et al., 2011).

7. Generalizations, Recent Developments, and Theoretical Insights

Modern research on VNS has advanced in several directions:

• Learning-Augmented VNS: Integration with deep reinforcement learning and transformer-based operator-selection modules, enabling dynamic adaptation to large, heterogeneous problems, and robust generalization to unseen instance types (2505.23098).
• Fitness Landscape Analysis: Theoretical studies elucidate the underlying fitness–distance correlation and search-graph structure (e.g., generalized hypercubes for multidimensional assignment problems), explaining the mechanics by which VNS escapes shallow or deep valleys and justifies the use of multi-start or grid-based starting solutions (Kammerdiner et al., 2021).
• Hybridization: VNS serves as a backbone for metaheuristics hybridized with exact solvers (e.g., Concorde/TSP, CPLEX/ILP), constructive heuristics, local search schemes, and stochastic acceptance mechanisms, extending its applicability to novel or highly-constrained combinatorial problems (Freitas et al., 2018, Martins, 2017).
• Problem Class Extensions: VNS has been successfully adapted to settings with complex feasibility spaces (e.g., multiple time windows, simultaneous clustering and feature selection, route-dependent constraints, generalized set partitioning), owing to its flexible neighborhood design and modular framework.

Theoretical limits on convergence rates, landscape connectivity, or guaranteed optimality remain problem-dependent, though empirical findings underline near-linear or polynomial scaling in many applied settings (Grbić et al., 2018, Kammerdiner et al., 2021).

---

References:

• "Variable neighborhood search for partitioning sparse biological networks into the maximum edge-weighted $N_k$0-plexes" (Grbić et al., 2018)
• "Variable Neighborhood Search for the Bin Packing Problem with Compatible Categories" (Santos et al., 2019)
• "Variable Neighborhood Search for solving Bandwidth Coloring Problem" (Matic et al., 2015)
• "BVNS para el problema del bosque generador k-etiquetado" (Consoli et al., 2015)
• "Landscape properties of the very large-scale and the variable neighborhood search metaheuristics for the multidimensional assignment problem" (Kammerdiner et al., 2021)
• "Randomised Variable Neighbourhood Search for Multi Objective Optimisation" (0809.0271)
• "Neigborhood Selection in Variable Neighborhood Search" (Geiger et al., 2011)
• "An intelligent extension of Variable Neighbourhood Search for labelling graph problems" (Consoli et al., 2015)
• "Exploring the Performance of Genetic Algorithm and Variable Neighborhood Search for Solving the Single Depot Multiple Set Orienteering Problem" (Kant et al., 2024)
• "Variable Neighborhood Search Algorithms for the multi-depot dial-a-ride problem with heterogeneous vehicles and users" (Detti et al., 2016)
• "A Variable Neighborhood Search for Flying Sidekick Traveling Salesman Problem" (Freitas et al., 2018)
• "Goal Clustering: VNS based heuristics" (Martins, 2017)
• "Learning to Search for Vehicle Routing with Multiple Time Windows" (2505.23098)