Dynamic Multi-Depot Vehicle Routing with Online Requests: Event-Driven Transformer--DRL and Rolling-Horizon Benchmarking
Event-driven learning and benchmarking framework for Dynamic Multi-Depot Vehicle Routing Problem with progressively revealed requests and evolving vehicle states Masked MLP and Transformer policies trained via behavior cloning and PPO, with deterministic feasibility masking preventing invalid vehicle-request assignments Nearest feasible heuristic outperformed all learned policies across routing quality, waiting time, stability, makespan, and runtime in a 20-scenario benchmark Rolling-horizon opt
Analysis
TL;DR
- Event-driven learning and benchmarking framework for Dynamic Multi-Depot Vehicle Routing Problem with progressively revealed requests and evolving vehicle states
- Masked MLP and Transformer policies trained via behavior cloning and PPO, with deterministic feasibility masking preventing invalid vehicle-request assignments
- Nearest feasible heuristic outperformed all learned policies across routing quality, waiting time, stability, makespan, and runtime in a 20-scenario benchmark
- Rolling-horizon optimization achieved lowest waiting times and makespan but at substantially higher computational cost
- Learned policies retained millisecond-level decision speed and generalized to instances with up to 80 requests without retraining, though no single method dominated all criteria
Why It Matters
This work bridges deep reinforcement learning and operations research for dynamic vehicle routing, a problem with direct applications in logistics, ride-sharing, and delivery services. It provides a rigorous benchmarking protocol that reveals the persistent competitiveness of classical heuristics against learned policies, offering practitioners a realistic assessment of when DRL-based approaches add value. The event-driven framework with fixed-prefix/flexible-suffix commitments also introduces a novel evaluation methodology for separating route reassignment from resequencing effects.
Technical Details
- Architectures: Masked MLP and Transformer policies trained through behavior cloning (BC) followed by proximal policy optimization (PPO), with deterministic feasibility masking to prevent invalid vehicle-request assignments
- Commitment mechanism: Fixed-prefix/flexible-suffix route commitments protect completed, active, and near-term decisions, enabling separate measurement of vehicle reassignment and resequencing costs
- Benchmarks: 20-scenario policy benchmark comparing learned policies against dynamic insertion heuristics (nearest feasible) and time-limited rolling-horizon optimization
- Scalability: Learned policies transferred to instances with up to 80 requests without retraining while maintaining millisecond-level inference speed
- Training variability: PPO had minimal average effect on MLP but improved Transformer performance on average, though with greater seed variability across five independent training runs
Industry Insight
- Classical heuristics like nearest feasible remain highly competitive against learned DRL policies in dynamic routing; practitioners should benchmark learned approaches against strong baselines before deployment
- The fixed-prefix/flexible-suffix evaluation protocol offers a practical framework for isolating different types of route disruption, which is valuable for real-world systems where decision stability affects operational costs
- For latency-sensitive applications requiring sub-second decisions at scale, learned policies offer transferability and speed advantages even when suboptimal in routing quality, suggesting hybrid approaches combining heuristic quality with DRL efficiency may be the most promising direction
Disclaimer: The above content is generated by AI and is for reference only.