A project me and a friend of mine from uni, @Phutoast, have been working on for some months now. The inspiration came from two seperate youtube videos: one about using convolutional neural nets in a reinforcement learning environment allowing an AI to learn how to play snake and another about autoencoders.
A problem with convolutional neural nets in reinforcement learning applications is the massive parameter count causing training to progress slowly. I wondered if it would be possible to - instead of tuning convnet parameters using neuroevolution - train the convnet seperately like the way an autoencoder is built. Then, once the autoencoder can reliably reconstruct input images, feed the lower dimension latent vector into an agent and train using neuroevolution off of that. I spoke to Phu and he pointed me to a paper on world models and OpenAI's evoluton strategies. We have a framework set up and are currently trying to figure out how to get better results out of the autoencoder.