Deep Learing to “teach” and AIhow to play videogames

One team of the University of Oxford to create a better performance of a AI in Atari videogames. They have inspired with the demostration of combine Q-learning with deep learning to play Atari games maked in the work of Mnih et al (2013).

They approach instead of making a better score to train the AI, they choose to “teach” the AI to imitate an expert gamer using Convolution layers in its architecture using the framework Theano. Because the trainng never use the score as an optimizing metrica and the AI is imitating a gamer, they named the method as Deep Apprenticeship Learning.

In order to process the data, this team preprocessed the frames to go from 128 color images to grayscale, resized the frames from 210 x 160 to 83 x 83, removing the background and finally normalizing it.

The results of the experiments are measured with the scores and, even the AI of D.A.L. method never knows its score, it performs better that Sarsa, the other AI.

Link: http://www.cs.ox.ac.uk/people/nando.defreitas/publications/apprenticeship.pdf