Sea Lock: A Multi-Agent Reinforcement Learning Framework for Adversarial Naval Blockade Simulation and Predictive Command and Control

Simulation-based operational analysis is only as good as the opposition it is tested against, and in most defence studies the adversary is scripted, so a course of action is tested against an assumption rather than an adversary. Faster sensing and faster execution sit upstream of that question; they do not answer it.

We present Sea Lock, a partially observable stochastic game of naval blockade over heterogeneous surface vessels in which manoeuvre, detection under electronic warfare and salvo combat are coupled inside one model, so two-sided interaction and uncertainty are built in rather than added afterwards.

Shared-policy PPO with finite-pool alternating self-play produces adaptive adversaries: across three independent seeds every run passes through an empirical balance window even though aggregate winners differ sharply, and a 7×7 frozen-checkpoint cross-play matrix shows strength is non-monotone in training iteration. A recurrent state-space model trained only on synthetic self-play experience predicts held-out engagement evolution with 4–15× lower open-loop error than persistence, within the recorded policy envelope. A typed event stream, an operator dashboard and a bounded language-shaped reward term are integration demonstrators, not primary claims.

Sea Lock is the maritime rendering of a broader adversarial- simulation engine designed for cross-domain composition. The same engine architecture extends to land logistics, air swarm operations, cyber resilience, and integrated multi-domain C2 — each rendering shares the engine while varying observation and action spaces, reward composition, and domain-specific physics. Digital Twin is one application; predictive wargaming, force structure analysis, doctrine red-teaming, and operator decision support are others.