pytorch-route-choice

PyTorch implementations of route choice models for transportation networks. Sparse operations via PyTorch Geometric and implicit differentiation via torchdeq.

Models

• RecursiveLogitRouteChoice (aliased as MarkovRouteChoice): link-based network route choice (Fosgerau et al., 2013), can also implement Maximum Entropy IRL (Ziebart et al., 2008).
• NestedRecursiveLogitRouteChoice: nested recursive logit with link-specific scale parameters (Mai et al., 2015).
• PerturbedUtilityRouteChoice: perturbed utility route choice (Fosgerau et al., 2022).

Installation

pip install git+https://github.com/ben-hudson/pytorch-route-choice

Examples

Example	Description
recursive_logit_example.py	Learn edge rewards from sampled paths via maximum likelihood
maxent_irl_example.py	Maximum Entropy IRL by matching observed feature counts
purc_example.py	Learn utility rates from observed edge flows
markov_traffic_example.py	Traffic equilibrium on the Sioux Falls network using deep equilibrium models

Sioux Falls link costs according to the Markovian Traffic Equilibrium model.

Development

pip install -e .
pytest tests

Architecture Notes

• The core computation solves linear fixed-point problems x = Ax + b via message passing on sparse graphs, using torchdeq for implicit differentiation.
• Graph data uses PyTorch Geometric's edge_index (COO format) convention: [2, num_edges] tensor.
• MarkovRouteChoice is aliased to RecursiveLogitRouteChoice.
• RecursiveLogitRouteChoice requires batch dimensions for inputs (use .unsqueeze(0) for single instances).
• node_dim is specified as a negative offset (default -1) to support batched operations.