This repo tries to reproduce the results of the following paper in PyTorch

• Gal, Y., McAllister, R. and Rasmussen, C.E., 2016, April. Improving PILCO with Bayesian neural network dynamics models. In Data-Efficient Machine Learning workshop, ICML.

However, it is failed to learn a good controller. The purpose of releasing this repo is to encourage if someone interested in to help to seek for potential bugs or check for reproducibility of the algorithm.

Env:

To be consistent with the original paper, I have modified the environment 'CartPole-SwingUp' with same 'physics' coefficients used in the paper.

Previously I have posted a discussion in PyTorch, with a visualization of computational graph of DeepPILCO.

Any help or comments are highly welcomed to make this algorithm work or check for reproducibility together.