Diverse Traveling Salesman Problem

• D-TSP is a variant of the Traveling Salesman Problem that computes sets of distinct tours meeting quality bounds while maximizing diversity for robust logistical applications.
• It leverages evolutionary and neural algorithms—such as (μ+1)-EA and Graph Pointer Networks—to optimize diversity metrics like entropy and Jaccard similarity.
• Adaptive strategies balance tour quality and diversity, achieving significant runtime gains and yielding practical benefits in fault tolerance and multi-agent planning.

The Diverse Traveling Salesman Problem (D-TSP) extends the classical TSP by seeking sets of distinct tours that simultaneously satisfy quality constraints—such as an upper bound on tour length or cost—while maximizing population-level diversity according to specified metrics. This framework is motivated by practical needs in logistics, fault-tolerance, and decision-making contexts where multiple, structurally distinct high-quality solutions are preferred over a single optimum.

1. Formal Definition and Diversity Measures

D-TSP formulations require, given a complete undirected graph $G=(V,E)$ with $|V|=n$ and cost function $d: V \times V \to \mathbb{R}_{\geq 0}$, the computation of a set $P=\{T_1,\dots,T_\mu\}$ of tours such that:

• Each tour $T_i$ satisfies $C(T_i) \leq B$ for some threshold $B=(1+\alpha)\cdot \mathrm{OPT}$.
• The diversity of set $P$, quantified by metrics such as edge-based distance or entropy, is maximized.

Canonical diversity measures include:

• Edge-based distance: $dist(T_i,T_j) = 1 - |E(T_i) \cap E(T_j)| / n$ (Do et al., 2022).
• Sum–sum diversity ($D_1$): averaged pairwise distance: $|V|=n$0.
• Sum–min diversity ($|V|=n$1): $|V|=n$2.
• Jaccard similarity: $|V|=n$3 (Yang et al., 3 Jan 2026).
• Entropy-based diversity: $|V|=n$4, where $|V|=n$5 counts edge appearances across population (Nikfarjam et al., 2021).

The computational problem is NP-hard, even when all feasible tours are known (dispersion problem) (Do et al., 2022).

2. Algorithmic Frameworks for D-TSP

Multiple evolutionary and neural frameworks exist for solving D-TSP under various diversity constraints:

Evolutionary Diversity Optimization (EDO)

• (μ+1)-EA: Maintains a population of $|V|=n$6 tours, mutates via local search (2-opt, 3-opt, 4-opt), and applies diversity-based survivor selection (Do et al., 2020).
• EAX-Based EDO: Uses Edge Assembly Crossover (EAX) with modifications to maximize entropy during subtour recombination, enhancing edge diversity and robustness (Nikfarjam et al., 2021).
• Niching Memetic Algorithms (NMA): Employs adaptive grouping (speciation) based on pairwise edge-distances, randomized first-improvement local search, migration, crossover (PMX), mutations, and group-wise elitist selection. Stage 1 finds feasible diverse seeds; Stage 2 refines via (μ+1)-EA optimizing $|V|=n$7 or $|V|=n$8 (Do et al., 2022).

Neural and RL-Based Methods

• Graph Pointer Network (GPN)+Dispersion: Autoregressively samples tours using a GCN+LSTM encoder-decoder augmented with sequence entropy loss to control diversity; a greedy 2-approximation dispersion algorithm selects the maximally spread subset (Yang et al., 3 Jan 2026).
• RF-MA3S (Relativization Filter–Multi-Attentive Adaptive Active Search): Pointer network encoder with RF for affine invariance and D parallel decoders. Adaptive active search optimizes joint tour quality and diversity, switching between optimality and diversity-driven updates (Li et al., 1 Jan 2025).

3. Theoretical Insights and Computational Complexity

The computational cost and theoretical bounds are determined by the choice of diversity constraint and optimization strategy:

• Dispersion Algorithm Complexity: Greedy dispersion for selecting $|V|=n$9 maximally spread tours from $d: V \times V \to \mathbb{R}_{\geq 0}$0 candidates runs in $d: V \times V \to \mathbb{R}_{\geq 0}$1 (Yang et al., 3 Jan 2026); 2-approximation guarantees for max-min dispersion hold in metric spaces.
• NMA Complexity: Per iteration, dominated by $d: V \times V \to \mathbb{R}_{\geq 0}$2 neighbor checks for local search and $d: V \times V \to \mathbb{R}_{\geq 0}$3 for grouping (Do et al., 2022).
• RL-based methods: GPN empirical runtime scales near-linearly with $d: V \times V \to \mathbb{R}_{\geq 0}$4 under GPU parallelism (Yang et al., 3 Jan 2026).
• Price of Diversity (PoD): The minimum achievable cost blow-up of $d: V \times V \to \mathbb{R}_{\geq 0}$5 edge-disjoint TSP tours vs. single optimum is precisely bounded: for $d: V \times V \to \mathbb{R}_{\geq 0}$6, $d: V \times V \to \mathbb{R}_{\geq 0}$7 is $d: V \times V \to \mathbb{R}_{\geq 0}$8 in 1D metrics and $d: V \times V \to \mathbb{R}_{\geq 0}$9 in general metrics (Berg et al., 17 Jul 2025).

4. Diversity–Quality Trade-offs and Optimization Strategies

Trade-offs between solution quality and diversity are critical in D-TSP:

• Entropy Loss ($P=\{T_1,\dots,T_\mu\}$0 in GPN): Higher entropy coefficient increases diversity (lower Jaccard), at the expense of higher tour costs (Yang et al., 3 Jan 2026).
• Active Search Switch ($P=\{T_1,\dots,T_\mu\}$1 in RF-MA3S): Smaller thresholds delay diversity-phase, favoring optimality; larger $P=\{T_1,\dots,T_\mu\}$2 accelerates diversity but can degrade average quality (Li et al., 1 Jan 2025).
• Speciation–Clustering: Niching mitigates feasible-region fragmentation by covering multiple basins but can induce within-cluster similarity, lowering minimum pairwise diversity (Do et al., 2022).

Empirical results on TSPLIB and large instances confirm that hybrid approaches (NMA+EA, GPN+dispersion) achieve higher diversity—even under tight quality constraints—than older single-phase or naive heuristics, with orders-of-magnitude runtime reductions on large problems (Yang et al., 3 Jan 2026, Do et al., 2022, Nikfarjam et al., 2021).

5. Experimental Benchmarks and Empirical Findings

Representative experimental setups:

Instance	Population (μ, k)	Quality Bound (α, c)	Best Achieved Diversity
berlin52	30	c=4	GPN-Tree: Jaccard 0.015 (Yang et al., 3 Jan 2026)
eil101	60	c=4	GPN-Tree: Jaccard 0.016 (Yang et al., 3 Jan 2026)
rat783	480	c=16	GPN-TreeM: Jaccard 0.002 (Yang et al., 3 Jan 2026)
TSPLIB51-101	up to 100	α=0.05–0.5	EAX-EDO Entropy-max: $P=\{T_1,\dots,T_\mu\}$3 maximized (Nikfarjam et al., 2021)

Empirical conclusions:

• GPU-accelerated neural frameworks (GPN) achieve diversity values superior to classical NMA and RL-MA3S, with 30–360× runtime improvements (Yang et al., 3 Jan 2026).
• EAX-EDO single-stage maximizes entropy and robustness: when edges of the optimal tour are banned, alternative high-quality tours are almost always found (Nikfarjam et al., 2021).
• RF-MA3S outperforms both NMA and neural baselines in joint optimality-diversity indices (MSQI, DI), generalizing to CVRP and POI tours (Li et al., 1 Jan 2025).
• For large population sizes and tight constraints, speciation and 2-opt mutation maintain better diversity coverage than more aggressive 3/4-opt (Do et al., 2020).

6. Structural and Practical Implications

Structural results indicate that diversity requirements, such as edge-disjointness, impose provable limits on achievable tour costs: in arbitrary metrics, doubling the optimum is unavoidably necessary for two disjoint tours (Berg et al., 17 Jul 2025). Greedy block-tiled gadgets and segment-depth arguments underpin these bounds and constructive algorithms.

Practical implications include:

• Fault tolerance: Large, well-dispersed tour sets support robust decision-making under route failures (Nikfarjam et al., 2021).
• Multi-agent planning: Diverse tours are crucial for strategic patrolling, multi-robot logistics, or dynamic re-routing.
• Scalability: Neural sampling methods and greedy selection are currently the most efficient for very large $P=\{T_1,\dots,T_\mu\}$4.

Limitations include scaling of evolutionary grouping, potential clustering in niching-based EAs, and diversity-quality parameter tuning in neural approaches (Do et al., 2022, Yang et al., 3 Jan 2026). Extending these frameworks to asymmetric, time-window, or submodular diversity models is a major future direction (Yang et al., 3 Jan 2026).

7. Future Directions

Open research challenges focus on:

• Developing scalable and adaptive diversity-optimal algorithms for $P=\{T_1,\dots,T_\mu\}$5 tours, and for other combinatorial domains.
• Integrating alternative diversity criteria, e.g., determinantal-point-process or submodular dispersion objectives (Yang et al., 3 Jan 2026).
• Tightening approximation bounds and improving matching-network sample efficiency in RL-based methods.
• Analyzing “Price of Diversity” functions for various D-TSP generalizations, including the impact of weaker diversity constraints on cost blow-up (Berg et al., 17 Jul 2025).
• Combining hybrid local search strategies to balance sum-sum and sum-min diversity objectives under strict quality bounds.

D-TSP research demonstrates that maximizing population diversity under optimization constraints requires algorithmic techniques and theoretical analyses beyond those of single-solution TSP. The field increasingly leverages both evolutionary and neural frameworks, with diversity metrics such as edge-based distances, Jaccard similarity, and population entropy serving as the foundation for scalable and robust solution-set generation.