Multi-agent Tabular Algorithms¶

This section documents MASA's first multi-agent learning coordinators. They are implemented under masa/algorithms/tabular/multi_agent/ and operate on PettingZoo ParallelEnv instances that are wrapped with a MASA constrained Markov game (CMG).

The current algorithms use independent learning: every agent owns a separate instance of MASA's existing tabular QL learner. A lightweight coordinator collects a simultaneous joint action, steps the shared environment once, and sends each agent its own transition.

For the time being, there is no joint-action Q-table, parameter sharing, or centralized critic.