Development and Analysis of CNN-based Computer Players for Game 2048
Datum konání: 7. 11. 2019 10:15, Autor: Kristýna Vávrová
Zveme Vás na přednášku prof. Kiminori Matsuzakiho z Kochi University of Technology v Japonsku. Přednáška proběhne 7. 11. 2019 v učebně S11 v 10.15–11.15.
Anotace přednášky:
Game 2048 is a stochastic single-player game, and its "easy to learn, but hard to master" characteristic has attracted not only many people but also researchers of artificial intelligence. There have been several studies on the development of strong computer players for game 2048. Among them, the most successful approach is based on T-tuple networks trained by reinforcement learning. Whereas neural networks (NN) are expected to be a powerful tool for developing strong players, the current NN-based players were only as strong as high-level human players.
The speaker's group have studied NN-based computer players for game 2048, especially based on convolutional neural networks. We tackled two research questions so far. One is the relationship between the depth of networks and the performance of computer players. The other is on the effectiveness of applying supervised learning with many playlogs from existing strong computer players. We identified that the number of convolution layers is important, i.e., networks with three or more convolution layers perform much better than that with two convolution layers. We also inspected the reason by analyzing the inner working of our networks (namely, a white-box approach).
In this presentation, we will introduce the development and the analysis of computer players for game 2048 based on convolutional neural networks.
