Learning to Dial-a-Ride: A Deep Graph Reinforcement Learning Approach to the Electric Dial-a-Ride Problem

Urban mobility systems are transitioning toward electric, on-demand services, creating operational challenges for fleet management under energy and service-quality constraints. The Electric Dial-a-Ride Problem (E-DARP) extends the classical dial-a-ride problem by incorporating limited battery capacity and nonlinear charging dynamics, increasing computational complexity and limiting the scalability of exact methods for real-time use. This paper proposes a deep reinforcement learning approach based on an edge-centric graph neural network encoder and an attention-driven route construction policy. By operating directly on edge attributes such as travel time and energy consumption, the method captures non-Euclidean, asymmetric, and energy-dependent routing costs in real road networks. The learned policy jointly optimizes routing, charging, and service quality without relying on Euclidean assumptions or handcrafted heuristics. The approach is evaluated on two case studies using ride-sharing data from San Francisco. On benchmark instances, the method achieves solutions within 0.4% of best-known results while reducing computation times by orders of magnitude. A second case study considers large-scale instances with up to 250 request pairs, realistic energy models, and nonlinear charging. On these instances, the learned policy outperforms Adaptive Large Neighborhood Search (ALNS) by 9.5% in solution quality while achieving 100% service completion, with inference times under 10 seconds compared to hours for the metaheuristic. Finally, sensitivity analyses quantify the impact of battery capacity, fleet size, ride-sharing capacity, and reward weights, while robustness experiments show that deterministically trained policies generalize effectively under stochastic conditions.

💡 Research Summary

The paper tackles the Electric Dial‑a‑Ride Problem (E‑DARP), a highly complex variant of the classical dial‑a‑ride problem that incorporates electric‑vehicle constraints such as limited battery capacity, nonlinear charging dynamics, asymmetric travel times, and strict service‑quality requirements (time windows, maximum ride time, vehicle capacity). Recognizing that exact combinatorial methods (branch‑and‑cut, branch‑and‑price) scale poorly beyond modest instance sizes and cannot meet real‑time decision requirements, the authors propose a novel deep reinforcement learning (DRL) framework that directly operates on edge‑level information of a real road network.

The core of the method is the Graph Edge Attention Network (GREAT) encoder, which ingests edge attributes (travel time, distance, energy consumption) rather than node coordinates. This design naturally captures direction‑dependent costs and non‑Euclidean network structures that are typical for electric vehicle routing. The encoder’s edge‑wise attention propagates information between edges sharing endpoints, enabling the model to learn representations that respect asymmetric energy usage and charging considerations.

On top of the encoder, an attention‑based decoder constructs vehicle routes sequentially. At each step the policy selects the next node (pickup, delivery, or charging station) conditioned on the current state, which encodes vehicle locations, battery levels, passenger loads, and the set of unserved requests. Feasibility masking enforces hard constraints (time windows, battery feasibility, capacity) during decoding, preventing the policy from generating infeasible actions.

Training leverages the POMO (Parallel Optimistic Monte‑Carlo) paradigm: multiple solution trajectories are generated in parallel from different starting points within a single episode, and the learning objective maximizes the best trajectory among them. This multi‑start strategy mitigates the high variance typical of combinatorial RL and encourages exploration of diverse routing patterns. The reward function is defined as total profit (revenue from served requests) minus total operational cost (energy consumption, passenger waiting and ride times, charging penalties), with additional terms to encourage full service completion. A curriculum learning schedule gradually increases problem size, starting from small benchmark instances and scaling up to large realistic scenarios.

Empirical evaluation comprises two case studies using ride‑sharing data from San Francisco. In the first set of benchmark instances (30–50 requests) the learned policy attains gaps of 0.0 % to 0.4 % relative to best‑known solutions, while delivering speed‑ups ranging from 20× to over 7,000× compared with exact solvers. In the second, large‑scale study (up to 250 request pairs, 500 nodes) the DRL approach outperforms an Adaptive Large Neighborhood Search (ALNS) metaheuristic by 9.5 % in objective value, achieves 100 % service completion, and produces feasible routes in under 10 seconds—orders of magnitude faster than the hours required by ALNS.

A comprehensive sensitivity analysis examines the impact of battery capacity, fleet size, ride‑sharing capacity, and reward‑weight settings. Results show that increasing battery capacity by 10 % yields 6–15 % profit gains, while appropriate completion incentives are crucial for maintaining full coverage. Smaller neural architectures often match or surpass larger ones, indicating that the edge‑centric design is highly efficient. Robustness experiments demonstrate that policies trained deterministically on static data generalize well to stochastic environments with up to 10 % demand or travel‑time variability, exhibiting only minor performance degradation.

The authors conclude that edge‑centric graph neural networks combined with reinforcement learning provide a scalable, high‑quality, and real‑time capable solution for electric vehicle routing with passenger‑centric constraints. They acknowledge limitations such as the need for multi‑depot extensions, dynamic request handling, and richer battery degradation models, and suggest future work on policy interpretability, integration with traffic prediction, and deployment in autonomous fleet management platforms.