Evaluation of Pretraining Methods for Deep Reinforcement Learning

In recent years, Machine Learning research has made notable progress using Deep Learning methods. Deep Learning leverages deep convolutional neural networks to extract features from data, and has been able to reinstate interest in Reinforcement Learning, a Machine Learning method for modeling behaviour. It is well suited for game type environments and Deep Reinforcement Learning broke headlines teaching a computer to play Atari games only from pixel inputs. The Swedish Defence Agency is conducting research to incorporate Deep Reinforcement Learning for creating virtual actors in military training simulations. While shown great potential, Deep Reinforcement Learning comes with the major downsides of being computationally heavy and needing extensive amounts of data to work well. To increase efficiency, this thesis evaluates methods for pretraining Deep Reinforcement Learning models using both novel and state-of-the-art pretraining methods. Evaluation is done in Atari games, using recorded data from a demonstrator playing the games. Results indicate that pretraining a network to learn features can have benefits when using Deep Reinforcement Learning. It also indicate that pretraining on demonstrator behaviour confounds learning. The thesis lastly discusses issues with the methods and data used, and highlights potential ways for improvement