This directory contains codes for supervised learning of Bridge bidding.

The code is provided with a environment of Contract Bridge to create observation tensors in bridge_state.py, which is a Python reimplementation based on OpenSpiel implementation of Contract Bridge.

Environment setup

We use Python 3.8 and Pytorch 2.0 to make datasets and train the network.

Making datasets

You can run the file make_dataset.py to generate a dataset of Bridge bidding.

For default, we use the dataset played by WBridge5, which is released by the author of Human-Agent Cooperation in Bridge Bidding. You can download the dataset here.

You can also create your own dataset using the encoding of OpenSpiel, in short

• The cards are represented as rank * kNumSuits + suit.
• The calls are represented in sequence: Pass, Dbl, RDbl, 1C, 1D, 1H, 1S, etc, starting from 52.
• Each trajectory in dataset is a sequence of actions, where first 52 actions are dealing, starting from north and going clockwise. Then the bidding and playing follows.

By running make_dataset.py, you will get datasets in save_dir, which are python dicts of str, torch.Tensor.

Training

After making datasets, you can run train.py to train the network.

The arguments are explained below:

• dataset_dir: the directory to your dataset made.
• save_dir: the directory to save trained network.
• learning_rate: the learning rate of optimizer.
• train_batch_size: the batch size for train set.
• valid_batch_size: the batch size for valid set.
• num_episodes: the number of iterations.
• eval_freq: the frequency of evaluation (how many iterations).
• device: the device to train and evaluate on.