Can't get it to work on Breakout

[Kokkini]

I tried using DQfD for Atari Breakout but it didn't work (after 4 millions training steps, the average reward reached 40 and doesn't increase pass that. The normal DQN can go over 100). I tried different exploration epsilon (0.1 and 0.01). Different batch size (32 and 64). Different replay buffer size (5e5 and 1e6). However, they couldn't go pass 40.
Do you know what I should do or do you have a set of hyper parameters that works for Breakout.

[ShoTakeuchi]

Thank you for the experiment and I've observed the same problem.
I'm sorry I don't know the exact cause.
Perhaps, This is because There are some bugs or the DQN implementation is not so efficient.
(I tested Original DQN Implementation(OpenAI baselines) this code is based on and it couldn't go over 100 after 4M steps.)
I'll continue to investigate the cause.

[Kokkini]

Thank you

[nianba]

any progress？