Q-learning
Reinforcement-learning method that learns action values (Q) for states; Deep Q-networks combine it with neural networks for high-dimensional inputs such as game pixels.
Q-learning is a reinforcement learning algorithm that estimates how good each action is in each state. An agent picks actions to maximize long-term reward while exploring alternatives. Classical tabular Q-learning works for small state spaces; neural networks can approximate Q-values for large inputs such as raw video frames.
DeepMind’s DQN (article) paired Q-learning with a convolutional neural network and stabilizers such as experience replay, showing human-level scores on several Atari games from pixels alone.