This repository contains the implementation for a paper (https://arxiv.org/pdf/1607.03290.pdf) that tackles the problem of bridge bidding without competition using Deep Q Learning. What is unique about this paper is how it feeds in the raw hands as one hot vectors, instead of hand crafted features. It reports an improvement over WBridge, the most successful Computer Bridge Player, in bidding.
The data used for training the model consists of a North and South hands and a measure of how good a bid is given these two hands. This is computed by taking an expectation over the East West hands by randomly distributing the remaining cards to the East and West five times and computing the average of the scores obtained from those 5 games. The scores are computed by double dummy, in the form of maximum tricks obtained and then finally converted to IMPs.
The script GeneratingData/run.sh generates the data in the required format and stores it in GeneratingData/data as a json. One hundred examples are given in this repository as a sample and more can be generated by modifying the main function in GeneratingData/DoubleDummy.cpp