About testing Deep Q-Network after training.Training data and test data do not correspond

Question

I have use https://github.com/keras-team/keras-io/blob/master/examples/rl/deep_q_network_breakout.py to play Atari Breakout.

Some of the latest training results： runn