Rainbow agent for Playing 2048 with ReLAx
rainbow_run.mp4
This repository contains an implementation of 2048 game which may be played manually in Jupyter and a custom Gym environment written on top of it.
Then, this environment was used to train Rainbow deep q-network (Rainbow DQN) agent.
Resulting actor shows a very solid performance casually achieving 2048 tile (52% of games), and rarely (~1% of games) achieving 4096 tile. Table of each tile frequency is shown below (100 games played):
| Tile Value | % Games Achieved |
|---|---|
| 2 | 100% |
| 4 | 100% |
| 8 | 100% |
| 16 | 100% |
| 32 | 100% |
| 64 | 100% |
| 128 | 100% |
| 256 | 99% |
| 512 | 98% |
| 1024 | 83% |
| 2048 | 52% |
| 4096 | 1% |
Training was run for 10m environment steps. The graph of average return vs environment step is shown below (logs done every 50k steps):
The distribution of estimated Q-values vs data Q-values for rewards-clipped environment is shown below:
