In this work, we present a deep learning model based on reinforcement learning that is tied to an AI agent. The agent successfully learns policies to control itself in a virtual game environment directly from high-dimensional sensory inputs. The model is a convolutional neural network, trained with a variant of the Q-learning algorithm, whose input is raw pixels and whose output is a Q-value directly associated with the best possible future action. We apply our method to a first-person shooting game - Doom. We find that it outperforms all previous approaches and also surpasses a human expert.

1. Sutton, R. S. & Bartro, A. G. (1998). Introduction to Reinforcement Learning. Cambridge, MA: MIT Press.
2. Bellman, R. E (1957). Dynamic Programming. , Princeton, New Jersey: Princeton University Press.
3. Bellemare M. G. , Naddaf, Y. , Veness, J. & Bowling M. (2013). The arcade learning environment: An evaluation platform for general agents. Journal of Artificial Intelligence Research, 47:253–279.
4. Bellemare, M. G. , Veness, J. & Bowling, M. (2012). Investigating contingency awareness using atari 2600 games. In AAAI.
5. Silver, D. (2016). Deep Reinforcement Learning. DeepMind Technologies.
6. Krizhevsky, A. , Sutskever, I. & Hinton, G. (2012). Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems 25, pages 1106–1114.
7. White, D. J. (1993, November). A Survey of Applications of Markov Decision Processes. The Journal of the Operational Research Society, Vol. 44, No. 11, pp. 1073-1096.

Machine Learning Reinforcement Learning Q-learning Dqn Cnn