Learning to Design City-scale Transit Routes

TL;DR

This work tackles the Transit Route Network Design Problem (TRNDP) by introducing an end-to-end reinforcement learning framework that builds city-scale transit networks sequentially using graph attention networks. A two-level reward scheme provides intermediate feedback during route construction and terminal feedback from full traffic simulations, enabling effective credit assignment over long horizons. The approach is validated on a real Bloomington, IN dataset with 143 nodes, 243 edges, census-based origin-destination demand, and 16 existing routes, showing substantial improvements over human-designed networks and traditional heuristics under multiple initializations and modal-split scenarios, including a notable increase in service rate and reductions in wait times. The authors release the dataset and code, demonstrating the potential of data-driven transit design and outlining future work to generalize to more cities and time-varying demand patterns.

Abstract

Designing efficient transit route networks is an NP-hard problem with exponentially large solution spaces that traditionally relies on manual planning processes. We present an end-to-end reinforcement learning (RL) framework based on graph attention networks for sequential transit network construction. To address the long-horizon credit assignment challenge, we introduce a two-level reward structure combining incremental topological feedback with simulation-based terminal rewards. We evaluate our approach on a new real-world dataset from Bloomington, Indiana with topologically accurate road networks, census-derived demand, and existing transit routes. Our learned policies substantially outperform existing designs and traditional heuristics across two initialization schemes and two modal-split scenarios. Under high transit adoption with transit center initialization, our approach achieves 25.6% higher service rates, 30.9\% shorter wait times, and 21.0% better bus utilization compared to the real-world network. Under mixed-mode conditions with random initialization, it delivers 68.8% higher route efficiency than demand coverage heuristics and 5.9% lower travel times than shortest path construction. These results demonstrate that end-to-end RL can design transit networks that substantially outperform both human-designed systems and hand-crafted heuristics on realistic city-scale benchmarks.

Figures (4)

• Figure 1: (a) The Bloomington transportation network with $143$ nodes and $243$ edges. (b) Spatial distribution of trip origins (blue) and destinations (red) showing concentrated activity areas with highest demand origin of $344$ veh/hr from node $1$ ($\blacklozenge$), and highest demand destination of $1,681$ veh/hr to node $129$ ($\bigstar$). (c) The $16$ transit routes currently operating in the city.
• Figure 2: Comparison of real-world and RL-designed transit route networks in the Bloomington network at modal-split parameter $\alpha=0.3$. Gray nodes and edges show the base network; colored elements indicate transit routes. The RL design utilizes $143$ unique edges with $39$ shared by multiple routes, spreading service more broadly with less redundancy than the real-world network which has $135$ edges, among which $47$ are shared. The marker indicates the transit center (node $96$).
• Figure 3: Comparison of transit network designs with random initialization at modal-split $\alpha=0.3$ and $\alpha=1.0$. The top four metrics reflect passenger experience, while the bottom three reflect operator performance. RL-designed networks achieve $68.8\%$ higher route efficiency than greedy demand coverage at $\alpha=0.3$ and deliver the shortest travel times ($41.53$ minutes). At $\alpha=1.0$, RL achieves the highest service rate ($74.5\%$) and $47.3\%$ higher route efficiency than shortest path. No single baseline dominates across all metrics, reflecting inherent trade-offs in multi-objective transit design. Results are averaged over $5$ independent designs.
• Figure 4: Overview of the policy network. A shared graph attention backbone maps $16$ node features, $2$ edge features, and edge connectivity to final node embeddings of size $64$. First, the node features are projected to $64$ channels. Then the representation passes through a stack of four GATv2 blocks in sequence with the output of attention heads averaged (instead of concatenated) to control channel growth. Each block uses pre-layer normalization, a GATv2 attention layer with dropout, and a residual path that copies the input when widths match or applies a linear projection when they differ. The actor head is an MLP with three hidden layers of widths $256$, $128$, and $64$ that produces a score per node, followed by a feasibility mask to obtain probabilities over valid nodes. The critic head concatenates global mean and global max pooling of the node embeddings, then feeds the result to an MLP with three hidden layers of widths $256$, $128$, and $64$ to predict a scalar value per graph.