Automatic Bridge Bidding Model Based on LSTM and Deep Reinforcement Learning

There are two phases for playing bridge: bidding and playing. We focus on the bidding phase, which is critical for information exchange. Yeh and Lin have designed a deep reinforcement learning bidding model, but this model has a significant non-competitive assumption -- the opponent always bids PASS. We remove this limitation to reflect actual bidding situations. Unlike previous methods in the literature, we use a Long Short-Term Memory (LSTM) model specialized for time series data combined with reinforcement learning, which allows the model to learn independently without relying on human bridge games for training. We also combine probability simulation and upper confidence bounds to choose bidding actions. Experiments show that our approach achieves an average cost of 4.56, outperforming Yeh and Lin's approach without the bidding limitation.