Super simple, no-frills A2C agent that achieves over 200 reward on Atari Breakout, with MG2033/A2C and openai/baselines as reference.
Here's the corresponding blog post.
To train a model from scratch, run
python a2c.py
I recommend a value of N = 50 or 100 for best results, though training does take some time with those values.
python a2c.py --n 100
Better graphs, Tensorboard visualizations, testing, and saved model files on the way.
| N | Max Reward | Iterations before overfit |
|---|---|---|
| 1 | ||
| 5 | ||
| 20 | 376 | |
| 50 | 397 | Less than 1020626 |
| 100 | 428 | Less than 1031646 |
| Inf |