We have implemented Gumbel MuZero with DJL (pure Java code running on top of PyTorch native) following DeepMind's MuZero paper with network improvements as suggested in DeepMind's MuZero Unplugged paper and the replacement of the maximizing over an upper confidence bound by DeepMind's Policy improvement by planning with Gumbel.
All the common logic is encapsulated in a platform module, while each game with its specific environment is implemented in a separate module:
- Two player zero-sum games with a final reward only:
- TicTacToe: The agent learns the perfect game in the tic-tac-toe integration test on a single GPU (NVIDIA GeForce RTX 4090). Perfect means that every possible decision in the decision tree is correct and stable from epoch to epoch. It also means that the agent is not exploitable in any way. However, it goes beyond exploitability: the agent chooses actions that are rewarded with the same probability in the optimal course of the game with the same probability. This means that the agent does not specialize, but remains broadly positioned. This ensures stability against potential changes in the environment.
- Go. We have started training the game of go, board sizes 5x5 and 9x9.
- One player games with a final reward only:
- PegSolitair: On the classic english board it learns perfect play: starting with one hole and end up with one peg in the middle.
We implemented onnx export for the muzero network.
You can find out the inference time while running MuZero on your edge device:
Install Java JDK 21.
We are using PyTorch 2.1.1 which is automatically installed during the build. However, it needs to have CUDA 12.1 and cuDNN 8.2.1 installed.
The gradle wrapper automatically downloads the correct gradle version. It is not necessary to install gradle. Just run the following command in the root directory of the project:
gradlew build
java -jar games/tictactoe/build/libs/tictactoe-0.7.0-SNAPSHOT-exec.jar
... more details on enpasos.ai
This project is licensed under the Apache-2.0 License.