Predicting human behavior in size-variant repeated games through deep convolutional neural networks

We present a novel deep convolutional neural network (DCNN) model for predicting human behavior in repeated games. The model is the first deep neural network presented on repeated games that is able to be trained on games with arbitrary size of payoff matrices. Our neural network takes the players' payoff matrices and the history of the play as input, and outputs the predicted action picked by the first player in the next round. To evaluate the model's performance, we apply it to some experimental games played by humans and measure the rate of correctly predicted actions. The results show that our model obtains an average prediction accuracy of about 63% across all the studied games, which is about 6% higher than the best average accuracy obtained by the baseline models in the literature.