This repository contains the code for a simplified implementation of "World Models" paper, (NeurIPS version) , where the controller is trained with A2C. Note that since we train both M and C together, we stopped gradients of the controller from flowing in to the world model, to allow for learning task independent features. Also the encoder is a simple fully connected network and will be converted into VAE and the model is currently a RNN with fully connected network, instead of mixture density network.
Results of this implementation on CartPole-v0 Open AI Gym environment. More results to follow.
