- Presentation
- Project Artchitecture
- Competition
- Baseline
- Environment & Setup
- Try it
- Fight it
- What I learned
Implementing deep reinforcement learning algorithms for the ICGA competition. This project is carried out for my 1st year of master internship at the LIP6 (Sorbonne University / CNRS).
AlphaZeroICGA/
├── src/
| ├── main/
| | ├── agents/ (Contains the jar files of the final agents)
| | ├── bin/ (Contains the binary files compiled from src_java)
| | ├── datasets/ (Contains the (state,distrib,value) datasets)
| | ├── final_model/ (Contains the final weights of the best models)
| | ├── libs/ (Contains the librairies such as JPY/Ludii...)
| | ├── models/ (Contains the current models)
| | ├── src_java/ (Contains all the source code in java)
| | ├── src_python/ (Contains all the source code in python)
| | | ├── brain/ (Contains the deep learning part)
| | | ├── mcts/ (Contains the vanilla MCTS and AlphaZero MCTS)
| | | ├── optimization/ (Contains the optimization part such as precomputations)
| | | ├── other/ (Contains utility files)
| | | ├── run/ (Contains files runned by java files such as dojo, trials...)
| | | ├── scripts/ (Contains all the scripts such as merge_datasets.py)
| | | ├── settings/ (Contains the hyperparameters and games settings)
| | | └── utils.py (File containing the utility functions)
| | ├── cluster_scripts/ (Contains the script to run alphazero on multiple nodes)
| | ├── cluster_logs/ (Contains all the logs output from the cluster scripts)
| | ├── alphazero.py (Script running the whole AlphaZero algorithm)
| | ├── alphazero_m.py (Same but only playing model vs model)
| | ├── alphazero_mm.py (Same but running on multiple nodes if on cluster)
| | ├── build.xml (Build file helping us run java commands, clean...)
| | └── notes.txt (Some notes I left while doing that project)
| └── test/ (Some Ludii tutorials and tests)
├── alphazero_env.yml (Conda environment save)
├── README.md
└── LICENSE
"The Ludii AI Competition involves general game playing events focussed on developing agents that can play a wide variety of board games. The events use the Ludii general game system to provide the necessary games and API. Games will be provided in the Ludii game description format (.lud). The version used for this competition (1.3.2) of Ludii includes over 1,000 games.
Three events are proposed :
- Kilothon: Best utility obtained on more than 1,000 games against UCT.
- General Game Playing (GGP): Competiton on games present or not in our library.
- Learning: A set of games are announced months before the actual competition, the agents are invited to learn before competing."
Here we focus on the learning event.
Links :
The different games of the learning event this year are :
- Bashni: https://ludii.games/details.php?keyword=Bashni
- Ploy: https://ludii.games/details.php?keyword=Ploy
- Quoridor: https://ludii.games/details.php?keyword=Quoridor
- Mini Wars: https://ludii.games/details.php?keyword=Mini%20Wars
- Plakoto: https://ludii.games/details.php?keyword=Plakoto
- Lotus: https://ludii.games/details.php?keyword=Lotus
We use deep reinforcement learning algorithms for this competition and we start with AlphaZero as a baseline. AlphaGo is an algorithm which can play Go at a super-human level using supervised learning and reinforcement learning. AlphaGo Zero can basically do the same but starting from scratch, hence the "Zero" in its name. AlphaZero does the same but it is able to play different games such as Chess and Shogi.
Links :
- https://www.nature.com/articles/nature16961 (AlphaGo)
- https://discovery.ucl.ac.uk/id/eprint/10045895/1/agz_unformatted_nature.pdf (AlphaGo Zero)
- https://arxiv.org/abs/1712.01815 (AlphaZero)
- https://arxiv.org/pdf/1903.08129.pdf (Hyper-parameters sweep on AlphaZero)
- https://www.scitepress.org/Papers/2021/102459/102459.pdf (Improvements to increase the efficiency of AlphaZero)
The games are hosted on the Ludii software, which is in java. Since we use python for our algorithms we will need a java-python bridge such as JPY. Microsoft Visual C++ 14.0 and Java JDK 7.0 are required to build JPY. We also need the Ludii software to run our algorithms in the environment. We compile a jar file in order to export our AI on Ludii thanks to ant so it is also required even though you can do it otherwise.
Links :
- https://github.com/Ludeme/LudiiPythonAI
- https://github.com/jpy-consortium/jpy
- https://visualstudio.microsoft.com/visual-cpp-build-tools/
- https://www.oracle.com/java/technologies/downloads/
- https://maven.apache.org/download.cgi
- https://ant.apache.org/bindownload.cgi
- https://ludii.games/download.php
First you need to clone Luddi and JPY repositories, then download C++ build and java JDK if you don't have it yet. Apache Maven is also required to build JPY. Once everything is installed go to the JPY folder and run :
SET VS100COMNTOOLS=<visual-studio-tools-folder>
SET JDK_HOME=<your-jdk-dir>
SET PATH=<maven-bin-dir>
python setup.py build maven bdist_wheel
If everything worked, you should have a build directory. Copy the content of the lib directory into the Ludii directory in a folder called /LudiiPythonAI/libs/. The Ludii jar file should also be moved to the libs directory. Finaly, you can build the jar file thanks to ant and the xml file, then export it in Ludii.
You might also have to specify some paths in the configuration files such as jpyconfig.py and jpyconfig.properties. You might aswell modify build.xml file in order to set the correct classpath for the JPY snapshot.
There is an alphazero_env.yml file which can be used to create a conda environnement from scratch with all the required librairies with the command conda env create -f alphazero_env.yml.
The required python librairies are :
- tensorflow-gpu (CUDA, cuDNN, TensorFlow...)
- onnx, onnxruntime, onnxruntime-gpu, tf2onnx
- numpy, matplotlib, keras
The only game which is trainable on a personal machine is the Connect Four game. The other ones can be trained only on CPU/GPU clusters.
Go to the src/main/ directory and run the next commands in a terminal :
nano src_python/settings/config.py : set the settings to run AlphaZero such as number of MCTS simulation, the model hyper-parameters...
nano src_python/settings/game_settings.py : set the settings for the different games AlphaZero can play such as number of rows, columns...
python3 alphazero.py <n_loop> <n_workers> : runs the whole loop (MCTS simulation with random moves -> dataset -> train model -> save model -> MCTS simulation with model predicting moves -> dataset -> ...). n_loop is the number of loop it will achieve. n_workers is the number of processes which will be executed in parallel.
python3 alphazero_m.py <n_loop> <n_workers> : runs the whole loop but not playing against MCTS vanilla, only playing model vs model.
python3 alphazero_mm.py <n_loop> <n_workers> <n_nodes> : same as alphazero_m.py but able to run on multiple SCAI GPU clusters nodes.
The python alphazero script does everything, the following commands are for debugging purposes :
ant clean : clean all the directories (bin/ build/ models/ datasets/).
ant build : compile the java file in bin/.
ant run_trials Dforce_vanilla=<force_vanilla> : runs the MCTS simulations only (randomly or using the model depending if there is a model in models/) and creates a dataset. force_vanilla is an argument which force the MCTS vanilla to play even if there is a model in models/.
ant run_dojos : runs a 1 versus 1 between the last model (the outsider) and the best current model (the champion model) and outputs some stats. If there is only one model (the champion), runs a 1 versus 1 between the champion model and the MCTS vanilla.
ant run_tests : runs tests against Ludii built-in AIs.
ant create_agent : takes the best model and build an agent as a jar file for the Ludii software.
python3 src_python/brain/train_model.py <learning_rate> <force_champion> : only trains the model using the dataset and saves the best model. force_champion is an argument which force the script to train de champion, default value is false.
python3 src_python/scripts/merge_datasets.py : merges all the datasets in datasets/ with an hash into a unique dataset.
python3 src_python/scripts/merge_txts.py : merges all the text files in models/ with an hash into a unique txt file.
python3 src_python/scripts/switch_model.py : switch outsider to champion and champion to old_star.
When the project will be over, the model will be available in the folder models/final_model/ and the Ludii AI will be in the folder agents/ as a jar file in order to load it in Ludii software. You will be able to load it against other AIs or against you on different games.
General knowledge :
- Papers implementation and understanding (AlphaGo, AlphaGo Zero, AlphaZero)
- Software architecture with different task communicating with each others (alphazero.py)
- Java wrapper for python with JPY
- Some additional knowledge about Linux system
Deep learning :
- Multi-headed neural networks (here for policy + value prediction)
- Huge CNN model with residual blocks and skip connection
- Features in the input matrix channels (such as player turn, number of moves played...)
Reinforcement learning :
- MCTS with UCB/PUCT scores
- State and action representation, reward system
- Temperature, dirichlet in policy for exploration etc...
- Using batch prediction to avoid wasting too much time in inference
- Time and memory optimization :
- Multithreading/Multiprocessing for the self play games
- Parallel execution of self-play games and dojos on different nodes of the cluster
- Usage of GPU clusters to train the models
- Code optimization because the algorithm is very time consuming (use of profilers)
- Precomputing functions which are called huge amount of time (in MCTS algorithm)
- ONNX format for faster inference with models