Replicating Results With dqn_atari.py

[gwen-dls]

I am trying to replicate the results in Table 9 of the appendix in this paper, however I am getting different results despite using the same parameters.

These are my results: (brackets denote standard deviation)
Game Mean Episodic Return
ALE/Enduro-v5: 214.8 (17.4)
ALE/SpaceInvaders-v5: 573.5 (8.9)
ALE/Pitfall-v5: -94.6 (12.1)

While in the paper they report

Enduro: 473.2 (22.3)
SpaceInvaders: 211.6 (14.8)
Pitfall: -24.9 (14.8)

I selected version 5 as it seemed to match most of their configurations, however some parameters don't seem to exist with this implementation of DQN such as:

Gradient momentum
Squared gradient momentum
Min squared gradient
Replay start size
Agent history length
Frame pooling (boolean)
No. of different colours

Any suggestions on how to address these parameters or if these have a different name within the code?