aerialblancaservices/v2rmp-neural-models

v2rmp Neural Models

Neural network models for the v2rmp route optimization engine. These are small, fast MLPs trained offline with PyTorch and exported to Candle-compatible safetensors for pure-Rust inference.

Models

Model	Architecture	Input	Output	Purpose
solver_selector	28 → 128 → 64 → 6	Instance features (28-D)	Solver probabilities (6 classes)	Recommend best VRP solver for an instance
quality_predictor	28 → 64 → 32 → 2	Instance features (28-D)	Gap (%), tour length (km)	Predict route quality before solving
automl	28 → 64 → 5	Instance features (28-D)	Max iter, temp, tabu, cooling, neighbourhood	Instance-aware hyperparameter tuning
move_scorer	16 → 32 → 16 → 1	Move features (16-D)	Improvement score	Score 2-opt candidate moves for local search
graph_embed	GraphSAGE placeholder	Node features (10-D)	64-D embeddings	Road network node embeddings (placeholder)

Feature Vector (28-D)

The input to solver_selector, quality_predictor, and automl is a 28-dimensional normalized feature vector extracted from a VRP instance:

• Geometric (8): n_stops, n_vehicles, avg_pairwise_km, lat_spread, lon_spread, density, area_km2, depot_centroid_dist
• Graph (8): knn_avg_degree, knn_max_degree, knn_clustering, knn_diameter, knn_mst_weight, knn_avg_shortest_path, knn_spectral_gap, knn_assortativity
• Demand (4): total_demand, demand_std, tight_capacity_flag, capacity_ratio
• Distance matrix (4): dist_mean, dist_std, dist_skewness, depot_dist_mean
• Objective (4): one-hot for min_distance, min_time, balance_load, min_vehicles

Solver Classes (6)

The solver_selector predicts among:

1. default
2. clarke_wright
3. sweep
4. or_opt
5. two_opt
6. neural_guided

Training

• Framework: PyTorch 2.x
• Loss: Focal loss (γ=2.5) + label smoothing (0.08) for classifier; MSE for regressors
• Optimizer: AdamW with cosine warm restarts (classifier) / ReduceLROnPlateau (regressors)
• Data: Synthetic VRP instances with solver distance labels
• Export: state_dict keys remapped to Candle lin{N}.weight / lin{N}.bias convention

Usage (Rust / Candle)

use candle_core::Device;
use candle_nn::{linear, Linear, Module, VarBuilder};

let vb = VarBuilder::from_tensors(tensors, DType::F32, &device);
let lin1 = linear(28, 128, vb.pp("lin1"))?;
let lin2 = linear(128, 64, vb.pp("lin2"))?;
let lin3 = linear(64, 6, vb.pp("lin3"))?;

License

MIT OR Apache-2.0 (same as v2rmp)

Generated by ML Intern

This model repository was generated by ML Intern, an agent for machine learning research and development on the Hugging Face Hub.

• Try ML Intern: https://smolagents-ml-intern.hf.space
• Source code: https://github.com/huggingface/ml-intern