README.md

deep-briscola

Tensorflow deep reinforcement learning agent playing Briscola card game.

What is Briscola??

This repository contains a Briscola game environment where different agents can play.

• RandomAgent: choose each move in a random fashion
• AIAgent: knows the rules and the strategies for winning the game
• DeepAgent: agent trained using deep reinforcement learning
• HumanAgent: yourself

Dependencies

sudo apt-get update && sudo apt-get install -y \
  python-dev \
  python3-pip
  
sudo pip install \
  tensorflow \
  hyperopt \
  matplotlib \
  pandas

Alternatively, a Dockerfile with all the dependencies installed is provided in this repo. To use it:

Install Docker on Ubuntu

$ bash docker/build.sh
$ bash docker/run.sh

Usage

Train a model

$ python3 train.py --network dqn --saved_model saved_model_dir

Play against trained deep agent

$ python3 human_vs_ai.py --network dqn --saved_model saved_model_dir

Play against AI Agent

$ python3 human_vs_ai.py

Features

Different networks implemented

Specify the network type using --network command line argument

• Deep Q Network
• Deep Recurrent Q Network
• WIP Synchronous Advantage Actor Critic (A2C)

Self Play

Train multiple agents using the self_train.py python script.

$ python3 self_train.py --network dqn --saved_model saved_model_dir

Results

• Training a Deep Q Network model for 75k epochs: achieved 85% winrate against a random player.

• Training a Deep Q Network model for 100k epochs using Self Play: achieved 90% winrate against a random player.