GeoHopNet: Neural Model for Urban UAV Logistics

• GeoHopNet is a neural combinatorial optimization architecture that integrates geometry-biased sparse spatial attention with modern Hopfield memory to address urban UAV site location and topology inference.
• It employs a distance-biased attention mechanism restricted to K-nearest neighbors, reducing computational complexity from O(N²) to O(NK) while scaling efficiently on large urban graphs.
• A variant applies low-rank matrix completion for reconstructing network geometry from partial observations, demonstrating robust recovery even with high data sparsity.

GeoHopNet is a neural combinatorial optimization architecture that integrates sparse spatial attention with modern Hopfield memory, designed specifically for solving dynamic site location problems for urban UAV (unmanned aerial vehicle) logistics at metropolitan scale. The framework addresses both the computational bottlenecks of standard attention models on large spatial graphs and the requirement for geometry-aware reasoning, yielding near real-time, high-quality solutions for multi-period facility planning problems. A separate, earlier line of research under the GeoHopNet name employs low-rank matrix completion over virtual coordinates and geodesic distances to reconstruct graph geometry from partial information, applicable to network topology inference and sensor network mapping.

1. Dynamic UAV Site Location Problem

The principal application context for GeoHopNet is the dynamic UAV site location problem. Given an urban graph $G = (V, E)$ of $N$ candidate facility sites, and a dynamic demand profile $\{w_i^t\}_{i,t}$ at nodes $V$ over multiple periods $t=1,\dots,T$, the goal is to select, at each period, $p_t$ open facility locations to minimize the total weighted travel cost incurred by serving demand points. The formulation is a mixed-integer program:

$$\min_{x,y}\;\sum_{t=1}^T\sum_{i\in V}\sum_{j\in V}w_i^t\,d(v_i,v_j)\,x_{ij}^t$$

subject to

$$\sum_{j}x_{ij}^t=1,\quad \sum_{j}y_j^t=p_t,\quad x_{ij}^t\le y_j^t,\quad x,y\in\{0,1\}$$

where $d(v_i, v_j)$ denotes the Euclidean distance. The problem is a computationally intractable instance of dynamic capacitated p-median, with severe scaling issues for classical DRL and combinatorial solvers once $N$ enters the urban (thousands) regime (Zhi et al., 14 Jul 2025).

2. Architecture: Distance-Biased Sparse Spatial Attention

GeoHopNet's encoder replaces canonical multi-head self-attention with an explicitly geometry-biased variant. For each attention head $N$0, the compatibility score between node embeddings $N$1 and $N$2 is

$N$3

where $N$4 are head-specific linear projections, $N$5 is a learned scalar, and $N$6 is a small multi-layer perceptron mapping absolute Euclidean distance into a learnable bias, thus integrating spatial information into the attention mechanism itself. This construction allows the network to encode geometric structure directly into relational reasoning.

Computational tractability is achieved by restricting, for each node $N$7, attention computation solely to its $N$8-nearest spatial neighbors $N$9. This transforms the usual quadratic complexity ($\{w_i^t\}_{i,t}$0) of pairwise attention into $\{w_i^t\}_{i,t}$1, rendering attention feasible for large urban graphs. For practical settings ($\{w_i^t\}_{i,t}$2), empirical runtimes scale nearly linearly with $\{w_i^t\}_{i,t}$3 ($\{w_i^t\}_{i,t}$4) (Zhi et al., 14 Jul 2025).

3. Modern Hopfield External Memory and Regularization

GeoHopNet incorporates a modern Hopfield network as external, differentiable memory in its decoder. The memory comprises keys $\{w_i^t\}_{i,t}$5 and values $\{w_i^t\}_{i,t}$6, supporting associative retrieval and content-based updates based on current decoder state $\{w_i^t\}_{i,t}$7:

• Read (retrieval): $\{w_i^t\}_{i,t}$8, output $\{w_i^t\}_{i,t}$9.
• Write: $V$0, modulated by a learned gate $V$1.

The mechanism corresponds to gradient descent on the Hopfield energy and admits a large number of high-capacity fixed points, enabling rich memory for combinatorial search.

To prevent collapse and promote representational diversity in memory slots, training is regularized by an orthogonality loss $V$2 and an entropy penalty on the attention weights $V$3, resulting in total loss:

$V$4

where $V$5 is the standard actor-critic RL objective (Zhi et al., 14 Jul 2025).

4. Performance Evaluation and Scalability

GeoHopNet has been empirically evaluated on synthetic dynamic p-median instances ranging from $V$6 to $V$7 nodes. Its performance was benchmarked against Gurobi, Teitz–Bart heuristic, Simulated Annealing, classic Attention Model (AM), and ADNet, with metrics including relative optimality gap and runtime.

Key results:

Instance Size (N)	Optimality Gap (%)	Runtime (s)	Comparison Baseline
100	0.21	0.051	ADNet: 0.27% gap in 0.093 s
1000	0.22	<0.1	ADNet: 1.11% gap in >1.0 s

GeoHopNet improves solution quality by 22.2% and is 1.8× faster than ADNet on N=100 instances, while for N=1000 it produces sub-percent gap solutions in real-time, where baselines fail or are orders of magnitude slower (Zhi et al., 14 Jul 2025). This suggests the model advances the practical boundary of solvable instances for dynamic site optimization in urban-scale logistics.

5. GeoHopNet for Topology Recovery from Partial Virtual Coordinates

An earlier incarnation of GeoHopNet focuses on network geometry recovery from partial information, particularly in contexts where physical coordinates are unavailable or inappropriate (Jayasumana et al., 2017). This approach leverages:

• Virtual Coordinates (VCs): Each node's coordinate vector is given by hop-distances to a randomly chosen subset of anchor nodes.
• Partial Geodesic Observation: Only a subset $V$8 of node-pair hop distances is observed (either directly or deduced from VCs).
• Low-Rank Matrix Completion: Reconstruction proceeds by solving

$V$9

where $t=1,\dots,T$0 is the projection onto observed entries and $t=1,\dots,T$1 is the nuclear norm. Solutions rely on singular value thresholding or alternating minimization. Under appropriate incoherence and randomness conditions, perfect recovery is possible with $t=1,\dots,T$2.

Empirical results on sensor networks and social networks show mean distortion $t=1,\dots,T$3 with up to 60% entries missing. The method generalizes to various VC systems and is robust to unknown or partial graph structures (Jayasumana et al., 2017).

6. Implications, Limitations, and Application Domains

GeoHopNet's architecture enables real-time, high-fidelity optimization and geometry recovery on large-scale spatial networks. In UAV logistics, it supports dynamic re-planning in the face of evolving urban demands and regulatory constraints. In network science, its matrix completion variant enables robust topology inference from limited observations.

A plausible implication is that the integration of geometric bias and memory-augmented attention constitutes a successful paradigm for scaling neural combinatorial solvers to previously out-of-reach urban and infrastructure optimization challenges. Limitations may include parameter selection for K-nearest neighbor sparsification and the challenge of hyperparameter tuning for memory regularization.

7. Relation to Broader Research Landscape

GeoHopNet’s design links together several methodological strands:

• Sparse attention mechanisms for scalable deep learning on graphs.
• Geometry-aware deep networks for spatial combinatorial optimization.
• Modern Hopfield memory architectures for combinatorial search.
• Low-rank completion for topology inference from partial geodesics.

These innovations extend the potential deployment of neural combinatorial solvers in logistics, sensor networks, and social sciences, addressing both data sparsity and computational scalability (Zhi et al., 14 Jul 2025, Jayasumana et al., 2017).