The World's Ports
Are Choking.

The Approach

Going beyond centralized control.

Traditional port optimization attempts to solve everything from a single point — one algorithm, one model, one decision-maker. This approach breaks down at scale. Real ports are too dynamic, too complex, and too distributed for any single system to manage effectively. When a centralized optimizer encounters unexpected conditions, the entire plan collapses.

CHMARL takes a fundamentally different path. Using cooperative hierarchical multi-agent reinforcement learning, the system operates on two tiers. At the strategic level, a high-level planner sets emission budgets, throughput targets, and fairness constraints using a primal-dual optimization framework that provably bounds cumulative emissions. At the operational level, each entity in the port — every vessel, every truck, every crane — is represented by its own AI agent that makes real-time decisions within those constraints.

These agents learn to cooperate, negotiate, and optimize collectively — much like a flock of birds that moves in perfect formation without a leader. A fairness-aware reward transformer with dynamically scheduled penalties ensures that no single operator is disadvantaged, achieving max-min cost equity across heterogeneous fleets.

And because the intelligence is distributed, the system is resilient. If one agent fails, the others adapt. If conditions change suddenly, the network recalibrates. The two-tier architecture enables linear scaling in agent count — meaning CHMARL can grow from a single terminal to an entire national port network without fundamental redesign.