This package implements the classic grid game 2048 for OpenAI gym environment.

pip install gym-2048

The package currently contains two environments

• Tiny2048-v0: A 2 x 2 grid game.
• 2048-v0: The standard 4 x 4 grid game.

• Observation: All observations are n x n numpy arrays representing the grid. The array is 0 for empty locations and numbered 2, 4, 8, ... wherever the tiles are placed.

• Actions: There are four actions defined by integers.

  • LEFT = 0
  • UP = 1
  • RIGHT = 2
  • DOWN = 3

• Reward: Reward is the total score obtained by merging any potential tiles for a given action. Score obtained by merging two tiles is simply the sum of values of those two tiles.

Currently only basic print rendering (mode='human') is supported. Graphic rendering support is coming soon.

Here is a sample rollout of the game which follows the same API as OpenAI gym.Env.

import gym_2048
import gym


if __name__ == '__main__':
  env = gym.make('2048-v0')
  env.seed(42)

  env.reset()
  env.render()

  done = False
  moves = 0
  while not done:
    action = env.np_random.choice(range(4), 1).item()
    next_state, reward, done, info = env.step(action)
    moves += 1

    print('Next Action: "{}"\n\nReward: {}'.format(
      gym_2048.Base2048Env.ACTION_STRING[action], reward))
    env.render()

  print('\nTotal Moves: {}'.format(moves))

NOTE: Top level import gym_2048 is needed to ensure registration with Gym.