The whole project is split into a Python and a Rust part. The Python part is responsible for training the neural nets, which will take only minutes, once there is training data. The Rust part is responsible for everything else - so both the engine itself, but also generating training data through rollouts based on older nets. Generating training data takes several days for each iteration.
The Rust code is split into 5 crates, displayed in the following diagram. Some (not all) external dependencies are also displayed, marked by "ext".
graph
coach ----> rayon("ext: rayon")
coach ---> engine
wildbg-c --> engine
wildbg-c --> logic
logic --> engine
web --> logic
web ---> engine
engine --> tract("ext: tract")
web ----> axum("ext: axum")
engine contains the core parts
- Move generation: given a position and a pair of dice – what are legal moves/positions following?
- Inputs generation: given a position – what are the proper inputs for the neural networks?
- Inference of the neural net: given the proper inputs – what are the cubeless money game probabilities for winning/losing normal/gammon/backgammon?
- Best move: Given a position and a pair of dice – what's the best move following?
For a more detailed documentation, see docs/dev/engine.md.
coach contains everything to generate training data and compare the performance of different neural networks.
This is achieved by finding many (more than 100,000) positions through self-play and subsequently rolling out these positions.
We then have a large set of positions and probabilities. In other words: we then have a large training set of inputs and outputs for the neural nets.
While engine only deals with cubeless equities, logic is handling cubeful equities, match play and cubes. This is mainly work in progress so far.
wildbg-c is a small layer on top of logic which allows C code to access wildbg.
For a more detailed documentation, see docs/user/wildbg-c.md.
web contains the HTTP JSON API for bots and GUIs to access the user facing features.
The OpenAPI documentation is autogenerated by web and can be accessed here: https://wildbg.shuttleapp.rs/swagger-ui/
Currently web depends on logic and engine. In the future we might clean it up and remove the dependency on engine.
benchmarks is not displayed in the diagram above.
It doesn't contain any regular source code but benchmarks of the other crates.
They can help you make performance sensitive changes.
You can run them by executing cargo bench -p benchmarks from the project's root folder.
For only one benchmark file, execute for example cargo bench -p benchmarks --bench rollout_bench.
Rollouts done by coach are parallelized with the help of rayon.
We don't implement neural nets ourselves. Instead, we rely on the ONNX format and use tract for inference.
web depends on several external crates, notably the web application framework axum.
The OpenAPI documentation is generated with the help of utoipa.
The training folder contains Python code for training the neural networks with PyTorch.
Training data first needs to be generated with coach. Existing training data can be found in the external repository https://github.com/carsten-wenderdel/wildbg-training
The Python part doesn't know anything about backgammon. Neural net inputs are generated by Rust code.
A detailed HowTo can be found in docs/dev/training.