Tanks! A multi-agent environment for ML-Agents

A multi-agent environment using Unity ML-Agents Toolkit where two agents compete in a 1vs1 tank fight game.

---

STATUS: Published, will have some minor updates.

---

The Environment

• Two agents compete in a 1 vs 1 tank fight game.
• The goal is to kill the opponent team while avoid being killed.

Observation Space

• Vector Observation space:
  • Bool canShoot (you can only shoot a snowball every 2 seconds).
• Uses a One-Hot Grid Sensor.

Agent Reward Function (dependant):

• (1 - accumulated time penalty): when you kill your opponent.
• (-1) When you're killed.

How to use it

• 1️⃣ Download the ml-agents Release 12 NOT ABOVE THIS VERSION (because we use a modified version of the gridsensor that's compatible with only MLAgents Extensions 0.0.1).
• 2️⃣ Clone this Tank repository.

On ML-Agents

• 3️⃣ Put the Tank.yaml file to your ML-Agents folder into .\config\ppo.

On Unity

• 4️ Open the Unity Project Project in ML-Agents repo.
• 5️ Download the package here
• 6️⃣ Import package > Custom package Tank.unitypackage file.
• Enable the built in package 'Particle System' and 'Audio' in the Package Manager if you have some Audio and Particle errors.

If you want to train

2 scenes:

• Train_NonRandomGeneration: Always the same environment (no random moving of the assets).
• Train_RandomGeneration: At each episode, the assets (trees, buildings) are placed randomly. 💡: A good idea, when you train, is to uncheck mesh render for all your assets, this way you can save computational power.

If you want to play on Unity

In scenes open Play scene.

You can also download the game on Itch.io.

Next Steps you can try

• Use MA-POCA, Multi Agent Posthumous Credit Assignment (a technique for cooperative behavior).