Repository of the master thesis BriscolaBot: Mastering Briscola with model-free Deep Reinforcement Learning.
We studied and implemented an AI Agent that plays Briscola, a famous italian card game.
You can play against our latest agent BriscolaBot-v3 both locally and on replit
- Click the following link https://replit.com/@LorenzoCavuoti/BriscolaBot
- Press Play
- Enter fullscreen mode
Copy paste the following in a bash terminal
git clone https://github.com/LetteraUnica/BriscolaBot.git;
cd BriscolaBot;
python3 -m pip install .;
python3 main.pyPuoi giocare contro il nostro ultimo agente BriscolaBot-v3 sia in locale che su replit.
- Clicca il seguente link https://replit.com/@LorenzoCavuoti/BriscolaBot
- Premy Play
- Entra in modalità a schermo intero
Copia e incolla il seguente codice in un terminale bash.
git clone https://github.com/LetteraUnica/BriscolaBot.git;
cd BriscolaBot;
python3 -m pip install .;
python3 main.pyIf you wish to contribute or have ideas on how to extend this project feel free to open an issue, reach out on Discord (username "lettera") or by email at lorenzocav97@gmail.com
The agent wins against average human players 59% of the time. More impressive are the results against expert players, where BriscolaBot-v3 has won 52% of the games, with only one expert player Silvio, being able to beat it with a score of 12-11.
The black dashed line represents the 50% win rate, while the gray vertical lines represent the 90% confidence interval.
All the human players we tested against expressed that the agent is very good at the game. One of the best performing players, Simone, found the agent frustrating to play against because it seemed to predict his hand cards.
The agent plays a total of 2048 games against 4 opponents sampled from the Agent Pool. The agent collects this experience and stores it in an Experience Buffer, which is then used to train the agent. After training with the PPO algorithm, a copy of the updated agent is inserted into the Agent Pool with frozen weights, and the process is repeated until the agent reaches the desired performance. For more information read the thesis or explore the notebook.
@misc{lorenzo2023briscolabot,
author = {Lorenzo, Cavuoti},
title = {BriscolaBot: Mastering Briscola with model-free Deep Reinforcement Learning},
year = {2023},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/LetteraUnica/BriscolaBot}}
}