PENGARUH BOBOT PADA DEEP CONVOLUTIONAL NEURAL NETWORK DENGAN SUPERVISED LEARNING UNTUK MENYELESAIKAN GAME 2048

2048 is a puzzle game that is played by sliding the tiles in a 4 x 4 board. The value of tiles of the board was 2n, i.e (2,4,6,16,…,32768). Every tiles can be slide into four directions (up, right, down or left), which every move bring out a new random tiles (between 2 or 4) at a random free space in the board. Deep Convolutional Neural Network (DCNN) with Supervised Learning has been used to learn and play 2048 by Kiminori Matsuzaki, et al. From the previous research, there is a conclusion to expansion the number of weights in the network architecture. Dataset used to learn the game 2048 as much 600.000.000 data, that acquired from 54,000 gameplays. This study focusing on enhancement of initialized weights to learn game 2048. The parameter of this study is a model that learnt four directions that will be used by DCNN to play the game. After increasing the number of weights, the level of accuracy of DCNN learning the game is 76.2%. The result model after learning phase were playing the game and get results ratio of 90.7% and 75 times completing 16384 tiles from 1,000 gameplays. DCNN is considered as the best method to classify and identify game 2048.