CoMaze Live Coding Session at EmeCom20

09.12.2020

We will be holding a live coding session at the NeurIPS workshop "Talking to Strangers: Zero-Shot Emergent Communication" on the interesting new multi-agent test bed CoMaze, inspired by a German board game called "Magic Maze"[1]

You can simply start playing the game by inviting friends to play with you. When creating a game, you can decide how many players are supposed to play. This creates a game-ID that you can share with your friends to join you. When entering the game you can decide which actions you want to control. The levels do not succeed automatically but you need to create a new game at a higher level.

Climb through the levels together and try to figure out, what the cryptic messages you send each other could possibly mean.

But, it will be much more interesting to develop AI that can play this game with a stranger. Many approaches to studying multi-agent cooperation and emergent communication rely on extensive amounts of shared training time. Our question is: Can we do that faster?

With this Colab notebook, you can try implementing your own player logic in the next_step() method. Simply uncomment one of the possible ways to start or join a game on the bottom and run the file. You can then join the game with the same ID through the web app and play together with your AI.

Join us on Dec 12th at 10:40-12:00 EST for a live coding session at NeurIPS!

Most important:
This is about having fun and testing some new ideas. We do not expect anyone to perfectly solve this game in the short amount of time given. Simply start with Level 1 and try to proceed from there. Hack some creative ideas about how you can infer meaning from the other player's utterances. Your agent instance will be stable through all levels with the same partner, so you can try to memorize things about this partner and use it at higher levels.

To participate, please register here with a unique team name. You will receive an individual upload link for your solution. We are looking forward to your submissions!

We prepared a more elaborated notebook for the session that will allow us to run a round-robin tournament with your agents. A simple rule-based agent is already implemented, which always moves to the closest unreached goal, thereby solving Level 1. Now it is your task to figure out your unknown partner's communication and use this knowledge to solve the higher levels.

Your implementation does NOT need to include reinforcement learning! We are open to any paradigms and ideas to solve this problem. This is especially meant to encourage new and unusual approaches!

However, please stick to the method structure as we present in this notebook to easily run our round-robin tournament. Only upload one .py file which contains a class MyAgent that inherits from the AbstractAgent class as it is in our notebook.

If possible, we will run the matches live in the lunch break of the workshop and present an analysis later on. The more creative your ideas on solving this problem are, the more material we have to discuss and bring the field forward.

[1] German Wikipedia Article on Magic Maze: https://de.wikipedia.org/wiki/Magic_Maze

Erstellen Sie Ihre Webseite gratis!