Data Science · Operations Research
Multi-Agent Optimization for Maritime & Fleet Route Scheduling Under Uncertainty
Historical Context: TICNES AI’s earliest production agent deployments in 2018 focused on multi-agent route scheduling for maritime and logistics operations — delivering 28% reductions in scheduling overhead and berth idle time. Today, we combine exact constraint solvers with natural-language dispatcher interfaces so human logistics coordinators can re-plan disrupted schedules in seconds.
Contents
1. Why static vehicle and vessel routing plans break within hours
In textbook Vehicle Routing Problems (VRP) and Berth Allocation Problems (BAP), travel times, customs clearance windows, crane availability, and fuel consumption rates are treated as known deterministic constants. An overnight batch solver computes an “optimal” schedule at 2:00 a.m. and hands a static PDF plan to dispatchers at 6:00 a.m.
By 9:30 a.m. in real-world logistics operations — whether managing coastal feeder vessels, port terminal slots, or regional distribution fleets — the static plan is obsolete:
- A vessel or heavy transport truck is delayed by weather or customs inspection, cascading into missed tidal or warehouse unloading windows.
- A priority cold-chain cargo container requires immediate re-routing before spoilage thresholds are breached.
- A road corridor or port crane goes offline for unscheduled maintenance, invalidating capacity assumptions across a dozen downstream routes.
2. Why LLMs should never solve combinatorial math directly
A common anti-pattern in modern AI projects is asking a Large Language Model to generate a route schedule directly from a prompt containing fifty delivery addresses and time windows. Autoregressive token prediction cannot solve NP-hard Mixed-Integer Linear Programming (MILP) or Constraint Programming (CP-SAT) problems; the model will inevitably violate capacity limits, invent travel times, or drop stops.
At TICNES AI, we enforce a strict division of labour between natural-language translation agents and mathematical optimization engines:
| System Layer | Technology | Responsibility |
|---|---|---|
| 1. Telemetry & Constraint Parser | Structured LLM Agent + AIS/GPS stream ingestion | Translates unstructured disruption notices (emails, port bulletins, dispatcher voice/text notes) into formal mathematical constraints |
| 2. Probabilistic Delay Forecaster | Gradient-boosted survival & time-series models | Predicts arrival-time distributions (`p50`, `p90`) conditioned on weather, port congestion and historical corridor delays |
| 3. Combinatorial Solver Core | CP-SAT / Adaptive Large Neighbourhood Search (ALNS) | Computes provably feasible, cost-minimising route re-allocations within a `5-second` warm-start budget |
| 4. Dispatcher Trade-Off Explainer | Grounded synthesis agent + visual Gantt diff | Presents the top 2 candidate re-schedules to the human dispatcher with explicit cost, SLA and fuel deltas |
3. Decentralised multi-agent negotiation for dynamic disruptions
When a localised disruption occurs (for example, a single truck breakdown or a two-hour berth delay), re-solving the entire global fleet schedule from scratch causes unnecessary “schedule nervousness” — changing assignments for forty unaffected drivers or vessels just to save 0.5% in theoretical distance.
Our multi-agent coordination protocol models each fleet unit or terminal resource as an agent with a local cost function and a stability penalty. When a disruption occurs, only the affected agent and its immediate spatial-temporal neighbours enter a bounded Contract Net auction to swap or absorb the delayed tasks, keeping the rest of the operational schedule untouched.
4. Explaining constraint trade-offs to human dispatchers
Dispatchers reject black-box optimization software when they cannot ask “Why did the system move Cargo B to Vessel 4?” or “What happens if we hold Truck 12 for forty minutes?”
By pairing our mathematical solver with a structured tool-calling agent, dispatchers can express ad-hoc constraints in plain Portuguese or English (“Bloqueia a rota da manhã para manutenção da grua 2 até às 14h e mostra o impacto nos clientes prioritários”). The agent translates the instruction into a solver constraint, runs a warm-started re-solve in under five seconds, and displays a visual before/after comparison awaiting human sign-off.
5. Measured operational gains
- 28% reduction in scheduling overhead and idle wait time validated across live logistics trials.
- < 5 seconds warm-start re-planning latency when responding to mid-day route or berth disruptions.
- Zero unverified autonomous dispatches: Every schedule commit requires explicit human dispatcher approval.