情報処理学会 第80回全国大会講演要旨

The accordance rate of haifu data is recognized as a benchmark for estimating the machine's learning ability. Traditional mahjong learning methods were mainly by human artificially extracting features and designing function blocks. Although there has been related researches in deep learning^[1] and CNN^[2] these two years, they still cannot exceed traditional methods' result due to their methods' limits. In this paper, based on previous CNN work, we built an improved data structure in order for not missing important information. For the deep neural network, we elaborately separate the information gained into different input parts and make the merge after some feature extraction. We show our result much better than the previous deep learning study, and also surpasses state-of-art traditional result^[3] on this task.