A 2048 deep ai that does not suck

solving 2048 with neural network(9 hidden layers). No search at all, just input the game grid(with one hot encoding), forward to the network and apply argmax on the outpust to select the action. live demo: https://tjwei.github.io/2048-NN/ In this video, 5 games played, all of them reach 2048. 4 of them reach 4096 and 2 reach 8192. Source code: https://github.com/tjwei/2048-NN The NN in this video is trained in supervised manner. A NN trained in reinforcement learning style is in https://www.youtube.com/watch?v=gzWaJVrlNxw