MARLA documentation

MARLA (Multi-Agent Reinforcement Learning Architecture) is a research platform for augmenting a reinforcement-learning offensive-security agent with an advisory, language-model-based Plan Maker.

Concretely: a recurrent-PPO agent plays NASimEmu scenarios directly. In the assisted variant, it can also learn when to pause and ask a frozen, schema-validated Plan Maker for advice on the current set of legal actions, then learns how much to trust that advice via a residual, per-step confidence-weighted adjustment to its own policy.

There is no “coordinator” agent and no shared blackboard: the RL Orchestrator drives the whole experiment lifecycle itself, and every piece of advisory traffic is validated by a Gatekeeper agent before it can affect a decision.

Note

This documentation describes the implementation as it exists in this repository.

Where to start

Indices and tables