On the use of Deep Autoencoders for Efficient Embedded Reinforcement Learning