Deepq algorithm can't converge when change to another atari game

[Ja1r0]

In the "deepq" document,three experiments are provided,including one atari game "Pong".The deep-Q algorithm works pretty good in "Pong",but when I change Pong to another atari pixel game "AirRaid",the algorithm can't converge.Specifically,I just change
env = gym.make("PongNoFrameskip-v4")
to
env = gym.make("AirRaidNoFrameskip-v4")
in "baselines/deepq/experiments/train_pong.py" without any other modification to the code and hyperparameters.And the algorithm can't converge even after 1400 episodes.I don't know why.

[cxxgtxy]

Good question. In deed, you should use train in experiment/atari dir

[zwfcrazy]

According to the original paper of DQN, DQN does not work well on games with very complex state space, such as the game Montezuma's Revenge. I'm not sure how complex Air Raid is, as I didn't find this game in the paper...
Maybe you can take a look how complex this game is...