MOASEI 2026 - AAMAS

Hello welcome to the MOASEI 2026 docs! MOASEI is a competition to see which policies can perform the best in open multi-agent systems. See competition details on the MOASEI 2026 website. Here we will discuss expectations on submissions, and the evaluation procedure.

Your objective is to construct a Agent class for your selected track which will perform the best across all environment configurations that we provide here Kaggle. On that kaggle there are 3 configurations for each track (domain). Wildfire (DW) and Cyber Security (CS) have stochastic transition functions which are seeded. We will run the one submitted policy for that track across all configurations shown, not shown configurations, across multiple seeds. Policies earn points according to which place they score in each configuration. Policies are awarded n-(k-1) points for n many participating policies and kth place. The policy with highest points in the track wins that track.

Instructions

We recommend you follow the installation guide to install free-range-zoo, then run one of the full quickstart (rideshare here) scripts to verify your installation, and see the basic usage guide to see a example of making an Agent.

Submission

You must submit the following:

1. The source code of your Agent class with a list of all dependencies.

2. The source code used to train/update your Agent.

3. A modified version of the code shown in evaluation which initializes, loads, and evaluates your model.

4. The learned weights of your Agent.