The main goal of this thesis is to create controller that is able to play games of Atari 2600 console. At first the algorithm Neural Fitted Q iteration is proposed as a model of this controller. Based on the experiments a model is changed to a policy-search method. The final solution is presented on several Atari games. The raw picture of game is reduced using deep networks and is transmitted to a recurrent neural network with LSTM neurons. This network decides which action is best to make. In the conclusion of the thesis is comparison of results with other similar articles.