WSC 2025 — Winter Simulation Conference
2nd place at WSC 2025 in Seattle with a multi-agent RL approach to container port logistics
Context
The Winter Simulation Conference's annual student competition challenges teams to model and optimize a complex operational system under tight constraints. Our team from Universidad del Valle de Guatemala competed on a container port logistics problem.
What we did
Modeled the container port operations using O2DES (an open-source discrete-event simulation framework) and combined it with a multi-agent reinforcement learning approach using MAPPO (Multi-Agent Proximal Policy Optimization). Bayesian optimization tuned the policy hyperparameters within the simulation budget.
How it works
The simulation models the port's stochastic arrivals, berth allocation, crane operations, and yard logistics. The MAPPO agents learn coordinated dispatching policies. Bayesian optimization sits one layer above, tuning the agents and the simulation parameters efficiently against expensive rollouts.
Outcomes
Second place at WSC 2025 in Seattle, against teams from established research programs. The most useful takeaway personally was tuning the boundary between simulation realism and optimization tractability — every modeling decision is a tradeoff, and the right level of abstraction depends on what you're optimizing for, not what the system "really" looks like.