One program that learned to play Atari games from the pixels
orig. “Playing Atari with Deep Reinforcement Learning” · Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, Martin Riedmiller
Most game-playing programs are told the rules. This one was not. It saw only the raw pixels and the score, then learned by trial and error which actions led to more points. It paired that trial-and-error approach, called reinforcement learning, with a deep network that read the screen. On several games it reached or passed human skill, all from the same setup with no game-specific tuning.
This was a striking proof that one general method could learn many different tasks from scratch. It launched a decade of work on agents that learn by doing, leading to systems that play Go, control robots, and tune other AI. It is the cleanest early example of learning a skill purely from feedback.
Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, Martin Riedmiller, DeepMind
Paste it and we'll explain it even more simply.