Why is dqn playing breakout so bad? I trained 400 epochs and the final result did not exceed 6 points.

[invisilmk]

I trained 400 epochs and the final result did not exceed 6 points. What parameters need to be adjusted?

[Trinkle23897]

Could you please provide your configuration? (e.g. network architecture, hyperparam, ...)

[invisilmk]

It was run on the original code, but increased to 400 epochs.

[Trinkle23897]

It's our fault that we have no time to provide the tuned code example in the examples folder currently.
A better solution is to change the example code's hyperparameter and network arch to the DQN paper's setting.
You can refer to #31 (comment) for more details.

[Trinkle23897]

@Mehooz

[invisilmk]

Thanks a lot.

[duburcqa]

@Trinkle23897 Closing ?

[Trinkle23897]

@Trinkle23897 Closing ?

Nope. It hasn’t been finished yet.

[davide97l]

I'm waiting for this too.

[Trinkle23897]

Here it is: we now provide a DQN script which can solve Pong in 1M environment step (about 2000 seconds wall time in my computer). #187
The main issue of the previous version is that it doesn't use the wrapper provided by Deepmind. This is quite crucial.
Many thanks to @davide97l