2048-dqn

An attempt to slove 2048 by dqn

Implementation of 2048 is based on ankitaggarwal011/2048-console, MIT license

Modified for better performance. The biggest changes is to use int instead of str internally.

Files

• 2048.py
  • Implementation of the game logic
  • Implementation of DQN using TensorFlow
  • Should work on both Python 2 and Python 3
• 2048-eval-http.py
  • A HTTP interface to utilize the trained model with other 2048 frontends
  • In Python 2
• 2048-eval-http-p3.py
  • 2048-eval-http, but in Python 3
• model.*.ckpt.*
  • Trained model

Performance

( As of last update to this doc )

model.1545678840.187.ckpt

Reaches about 20000 pts in average.

Biggest tile is 1024 most of the times, sometimes reaches 2048, small chance to reach 4096.

Design

Train fixed times after each episode instead of each step to reduce training time and increase stability

State

log2 of the table

Reward

log2 of the addition to the score

HTTP interface

json pretty-printed for readability, but can be minified in practice

Recieve POST with json representing the game table

{
    "rows": [
        {
            "columns":[
                {
                "val":0
                },
                {
                "val":0
                },
                {
                "val":0
                },
                {
                "val":0
                }
            ]
        },
        {
            "columns":[
                {
                "val":0
                },
                {
                "val":0
                },
                {
                "val":0
                },
                {
                "val":0
                }
            ]
        },
        {
            "columns":[
                {
                "val":0
                },
                {
                "val":0
                },
                {
                "val":0
                },
                {
                "val":2
                }
            ]
        },
        {
            "columns":[
                {
                "val":0
                },
                {
                "val":0
                },
                {
                "val":0
                },
                {
                "val":2
                }
            ]
        }
    ]
}

Response with an action

{
    "action":2
}

where 0, 1, 2, 3 are left, bottom, right, top respectively