Recently, neural networks achieved major breakthroughs in many areas, which had been unfeasable for classical algorithms, e.g. in speech and image recognition, as well as in game playing AI. To apply this concept to the so far unsolved German card game Doppelkopf, we implemented an AI that combines the classical upper confidence bounds applied to trees (UCT) algorithm with a special type of recurrent neural networks, the long short-term memory (LSTM). We hypothesized this combination to be a promising approach, as the large search space of Doppelkopf would be reduced by using additional information from this neural network, as done in Go or Poker, for instance.
As a result, our LSTM predicted the next card in a Doppelkopf game at an average human level, and thereby, improved the tree search observably in a sense that promising nodes were visited earlier. Despite the impressive prediction accuracy of the LSTM, the combined, UCT+LSTM player achieved a score of only indicating that the LSTM did not improve the UCT player significantly.
As an outlook we propose several possibilities to improve this approach.