An Introduction to Deep Q-Network (DQN) can be found here.

Q-learning learns the action-value function $Q(s, a)$: how good to take an action at a particular state. Q is called the action-value function.

DQN is an approximation of the classic Q-learning algorithm by a neural network.

Advantages of DQN:

  • Capable of handling high-dimensional sensory inputs (states). [1]

Disadvantages of DQN:


[1] Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A.A., Veness, J., Bellemare, M.G., Graves, A., Riedmiller, M., Fidjeland, A.K., Ostrovski, G. and Petersen, S., 2015. Human-level control through deep reinforcement learning. nature, 518(7540), pp.529-533.