AGTNEWS
A per shipment AI won on freight volume and delivery rate. A simpler rule won on recovery time. The most powerful AI in the study collapsed in deployment.
The first three days of the simulated hurricane knocked out the Gulf Coast rail interchange, flooded the Memphis hub's access road, and slowed trucks out of Atlanta to a crawl. At the Jacksonville inland port, a dispatcher had to decide, in real time, which of 15 freight hubs should take which of the next 200 shipments: by truck, by rail, or by a barge-and-rail intermodal handoff. The decision had to be made per shipment, not per day, and the network's capacity was being squeezed faster than any rule-of-thumb could absorb.
A new operations-research paper tested three routing policies against exactly this kind of stress, across 30 matched hurricane episodes on a 15-hub intermodal freight network. The result was not a winner. It was a tradeoff map, and a deployment failure the authors say is the most important finding of all.
The paper, "Per-Shipment Multi-Agent Reinforcement Learning for Intermodal Freight Routing Under Hurricane Disruption", frames the routing problem as a Dec-POMDP, a shared decision problem in which no single agent sees the whole network. The twist is the action granularity: the AI does not pick a daily policy for the whole hub; it makes a routing decision for every single shipment. That distinction matters because most freight-routing software optimizes a plan, not a queue, and the two behave very differently when a storm is rerouting traffic by the hour.
The first AI the team trained is IPPO, an independent-learner policy trained with a method called Centralized Training with Decentralized Execution, or CTDE. In plain terms: each routing agent trains with access to the full network state, but at deployment each agent only sees its own local view. The intuition is the same as a regional trucking manager who has studied the entire company playbook but, on shift, only sees the docks in front of her.
The second AI is MAPPO, a shared-critic variant. Instead of training each agent independently, MAPPO trains a single "critic" network that evaluates decisions across the whole system at once, a more powerful and more centralized approach that has produced headline-grabbing results in games and protein folding.
The third is not an AI at all. It is a capacity-aware heuristic, a rule-based policy that shifts load away from any hub whose remaining capacity falls below a threshold. It also has "privileged state access" in the benchmark: it can see the same network-wide information that the AI critics do.
Across 30 matched hurricane episodes, IPPO moved the most freight: throughput rose 12.7% and delivery rate improved over both baselines. The capacity-aware heuristic still won on the metric the authors call the Resilience Index, a measure of how fast a disrupted network recovers, and on average delay. No single policy dominated.
The most interesting finding emerges when the storm does not just disrupt the network but also floods it with demand. In a regime where total shipment volume was 2.9 times the available capacity, IPPO's lead on the Resilience Index grew to 6.4%. Read carefully, this is the opposite of the typical AI-benchmark story: the learned policy helps more when capacity is scarce, not when there is more data.
Then came the deployment failure. The MAPPO variant, the most powerful and most centralized AI in the study, collapsed. Its Resilience Index fell to 0.811, well below the heuristic baseline. The cause was a queue mismatch: the queue length the critic learned to expect during training did not match the queue length it actually encountered at deployment. Retraining recovered MAPPO's Resilience Index to 1.018, but IPPO still won on throughput, and the heuristic still won on delay. The centralized variant was not production-ready, and the gap was not about intelligence. It was about the brittle assumption that deployment looks like training.
That gap is the line a public reader should remember. The paper is a simulated 15-hub benchmark, not a deployed system, and "AI" here is not a chatbot. It is a routing policy. But the lesson travels: a model with a global view is not always the one that should run a critical-infrastructure system, especially when climate stress is changing the queue faster than training can keep up.
For a dispatcher, a logistics executive, or a public-works planner, the decision rule is concrete. Pick the policy that matches the metric your supply chain actually has to optimize. If throughput and on-time delivery are the constraint, IPPO is the better choice. If resilience and delay dominate the metrics that matter most when a storm is shutting down lanes, a simpler capacity-aware rule is the better choice, and the all-seeing AI is the one to keep on the bench until its critic stops assuming deployment looks like training.