원문정보
초록
영어
Nowadays, machine learning and deep learning has been becoming popular and useful in applying to resolve people’s problem. Specially, in HCI (Human Computer Interaction) field like robots or automatic game programs. Go-Game (the game of Go) is still a challenge in coding to get the wisest moves each turn to achieve the winner at the end of a game. In our work, we suggest the next move based on Convolutional Neural Networks (CNNs) and make evaluations and comparisons to gamers separate in 3 ranks (levels). We train 5-layers CNNs by supervised learning from a database of human games using the board-states. The network suggests the move of the selected player and the others player can be helped or not- depend on playing option. The program can also play the game automatically without human interactions during all the game progress (Machine-Machine game). In the other way, our program can interact with a human-player and accept move commands from player (Human-Machine or Human-Human). This technique allows Go-game program play the game without searching as traditional program but trained by convolutional neural networks. In our tests, we separate in 3 levels and use totally 598,472 board-states for training data. Our main aim is to help people who are the newbie in playing Go-game. With this technique, we expected that we can apply to develop AI programs and devices with more and more effects and higher performance.
목차
1. Introduction
2. Training Procedure and Architecture of CNN
3. Data
4. Results
5. Calculating Score Table
6. Discussion
References