AI for Predicting Human Behavior through the Example of the Board Game Dixit

초록

The goal of this study is to create an AI that can replicate and predict human behavior in board games, achieving human-level performance. As a case study, we used the board game Dixit, a card game that stimulates creativity and imagination. In Dixit, players, typically three or more, describe the cards in their hands to encourage others to guess which card they are referring to. We aimed to design an AI model for predicting digital cards and to identify the optimal performance through a comparison and analysis of various data preprocessing and machine learning techniques. In this study, we addressed the problem of predicting the card chosen by the storyteller (the person providing the description), using different embedding techniques (such as Word2Vec and GloVe) and LSTM models to improve prediction accuracy. The results showed differences in accuracy depending on the data preprocessing and modeling approaches. Notably, treating card numbers as words and inputting them into the model led to improved performance. Additionally, adding LSTM and Bidirectional layers to the model resulted in accuracy higher than human players.

키워드