Status: Not really active (not a lot of time to develop and support this repo, breaking changes may occur). The program is checked in the case of n_agents=2

DeepGame

Here you can see implementations of the following single- and multi-agent learning algorithms:

1. Vanilla Q-learning algorithm (Every agent behaves in optimal Q-value way)
2. Nash Q-learning (Agents try to achieve Nash-equilibrium)

Prerequisites:

• Python>=3.6
• Numpy>=1.14.1
• Nashpy>=0.0.17
• Matplotlib>=2.2.2

Also it is based on https://github.com/ml3705454/mapr2 repo and you have to install mapr2 module:

1. Clone rllrb

cd <installation_path_of_your_choice>
git clone https://github.com/rll/rllab.git
cd rllab
git checkout b3a28992eca103cab3cb58363dd7a4bb07f250a0
sudo pip3 install -e .

2. Intsall other dependencies

sudo pip3 install joblib,path.py,gtimer,theano,keras,tensorflow,gym, tensorflow_probability

3. Intsall maci

cd maci
sudo pip3 install -e .

Instructions for running the program

Run the file run_grid_game.py with the following command and default parameters:

python run_grid_game.py

If your want to change the size of the grid world, number of iterations or some hyperparametes you should look through these arguments of run_grid_game.py:

• --grid_size, the size of the grid world. Default one is 3,
• --gamma, the value of the Gamma in Bellman equation. Default one is 0.95,
• --epsilon, the size of the epsilon. Default one is 0.5,
• --iterations, the number of steps in the grid World. Default one is 1000,
• --learning_rate, the learning rate (alpha) value for Q-learning method. Default one is 0.9
• --method, the method to choose (Q-learning, Nash-Q). Default one is 'Q'