Farama-Foundation/MOMAland

What it solves

MOMAland provides a standardized way to develop and compare multi-objective multi-agent reinforcement learning (MOMARL) algorithms. It solves the lack of a common API and consistent benchmarks for environments where multiple agents must optimize for multiple, often competing, goals simultaneously.

How it works

It extends the PettingZoo API standard, providing a set of environments where rewards are returned as vectorized numpy arrays instead of single scalar values. This allows learning algorithms to handle multiple objectives. The library also includes a LinearizeReward wrapper to convert these vector rewards back into scalars using weights if a user wants to shift to single-objective multi-agent RL.

Who it’s for

Researchers and developers working on multi-objective reinforcement learning and multi-agent systems who need a consistent framework for benchmarking and algorithm development.

Highlights

• Standardized API: Built on PettingZoo standards for seamless communication between algorithms and environments.
• Vectorized Rewards: Supports environments that return multiple reward objectives per agent.
• Built-in Benchmarks: Includes environments from MOMARL literature and multi-objective versions of classical environments like SISL or Butterfly.
• Integrated Algorithms: Provides a set of self-contained learning algorithms (e.g., MOMAPPO) implemented in a single-file style for clarity.