This page contains current information and data about the PePPEr algorithm. See also this website.

PePPEr stands for 'Policy Search with Probabilistic Principal Component Exploration', in Latex we add an additional C with PePP_CEr. The algorithm combines dimensionality reduction and policy search, such that the low dimensional manifold can be adapted during the learning process. Thus a high and a low dimensional probability distribution are combined to get a strong exploration along the manifold.

The algorithm was presented on the IROS 2014 in the paper 'Latent Space Policy Search for Robotics'.

You can find a test implementation of the algorithm here. However, these implementation is not runtime optimized but may support the understanding of the algorithm.

For further question, please feel free to contact Kevin Luck.