🎲 An AI agent that using Monte Carlo Tree Search to master the tile-laying game of Azul.
This repository contains a framework to support developing autonomous agents for the boardgame AZUL, published by Plan B Games. The game frame is forked from Michelle Blom's repository, and GUI is developed by Guang Hu and Ruihan Zhang. The AI agents are developed by Aaron Yau, Alan Ung and Aoqi Zuo. The code is in Python 3. This framework is able to run AZUL with two agents in a 1v1 game, and GUI will allow user to navigate through recorded game states. In addition, a replay can be saved and played for debug purpose.
Some information about the game:
- https://en.wikipedia.org/wiki/Azul_(board_game)
- https://boardgamegeek.com/boardgame/230802/azul
- https://www.planbgames.com/en/news/azul-c16.html
- https://github.com/michelleblom/AZUL
We ranked 6th in the last tournament and 2nd in the final playoff. Our total score in the tournament is the best. However, our win rate is the worst in the first eight. In the final playoff, we beat 404Error (who ranked first in the final tournament).
| Position | Team | Win | Tie | Lost | TotalGame | TotalScore | FAILED | FinalScore |
|---|---|---|---|---|---|---|---|---|
| 1 | 404Error | 240 | 1 | 27 | 268 | 14755 | 0 | 26775 |
| 2 | AlphaAzul | 230 | 4 | 34 | 268 | 14011 | 0 | 25591 |
| 3 | Amazing_Azul | 192 | 3 | 73 | 268 | 14703 | 0 | 24363 |
| 4 | StaffTeamEasy | 201 | 3 | 64 | 268 | 14163 | 0 | 24273 |
| 5 | thisisgroup1 | 180 | 4 | 84 | 268 | 14692 | 0 | 23772 |
| 6 | Diamond_Three | 156 | 8 | 104 | 268 | 15466 | 0 | 23426 |
| 7 | UnionOfThreeCities | 212 | 6 | 50 | 268 | 12662 | 0 | 23382 |
| 8 | AutoChess_Queen | 203 | 7 | 58 | 268 | 13055 | 0 | 23345 |
We also provide a demo for the project. which can be find here:
https://www.youtube.com/watch?v=4tPvgutIbvI
The code uses three library that required to be installed: numpy, func_timeout, tqdm, which can be done with the following commands:
pip3 install -r requirements.txt or
pip install numpy tqdm func_timeoutIf have both python 2 and python 3 installed, you might need to use following command:
pip3 install numpy tqdm func_timeoutTo run our best agent with the naive player:
python runner.py -r Diamond_Three.myPlayer -b naive_player
We also provide some other agents, you can find them in players/Diamond_Three/.
The code example can be run with command:
python runner.py, which will run the game with two default players (naive_player).
A standard example to run the code would be:
python runner.py -r naive_player -b random_player -s Other command line option can be viewed with argument: -h or --help
When a game ends, the GUI will pause to allow user selecting each states on listbox (right side of the window), and it will change accordingly. And replay file will be generated once the GUI window is closed.
For Debug purpose: Please use the Example.ipynb to start
Extra function
- timeout limit
- timeout warning and fail
- replay system
- GUI displayer (allow switch)
- delay time setting
class and parameters
AdvancedRunner
Runner with timelimit, timeout warnings, displayer, replay system. It returns a replay file.
ReplayRunner
Use replay file to unfold a replay
GUIGameDisplayer
GUI game displayer, you coud click items in the list box and use arrow keys to select move.
Check our assignment report for more details: report
We very much welcome your pull request if you improve our agent. The following ideas are some aspect we think can improve the result our agent.
- Use some heuristic functions like the one here https://cmath-school.com/blog/azul-ai to do pruning or make a more accurate future reward function.
- We leave many
TODOin our code, where we think can be improved in the future. - Actually, you can also reuse the subtree generated by the previous round and continue update the subtree to gain a better result. We didn't implement that because we did not fully understand the algorithm when implementing it.
Please read CONTRIBUTING.md for details on our code of conduct, and the process for submitting pull requests to us.