—
- env
- —
- action_space
- —
- observation_space
- —
What a DQN agent would see
t−3t−2t−1t
DeepMind's pipeline, reproduced live: the element-wise max of the last two frames (the 2600 flickered sprites on alternate frames), converted to luminance and resized to 84×84, with the last four stacked as the network input. JEV does not read this — it reads the structured state on the right — but this is the observation the published Atari agents were trained on.