I trained an AI to play 2048 using deep learning.
Deep 2048 plays like a human (well, me), because it’s not a brute force approach to playing the game 2048.
It had learned to play by observing my gameplay for just 20 games, and it had already picked up the basic concept of how the game works.
If you play the video at quarter speed, you’ll notice that the AI had learn some pretty interesting trick that I use when playing 2048.
Technical Details
The AI is a deep neural net that consists of 2 convolutional layers, 64 & 96 features respectively, and 2 fully connected layers with 512 neurons on the hidden layer and 4 outputs on the last layer.
The neural net takes the board configuration as input (16 inputs) and output a probability distribution for each direction (up, down, left, right).
The AI is trained via supervised learning. The surprising thing is how well the AI is playing with training data from just 10 games. And, the AI above is trained only with 20 gameplay.
