Abstract

Speaker 1 – Gregor Simm Title – Exploring Chemical Reaction Networks with Gaussian Processes Abstract – For the theoretical understanding of chemical systems, many costly quantum chemical calculations have to be performed. Therefore, approximate methods are often employed which suffer from low accuracy. We address this issue by applying Gaussian processes to learn the error of these approximate methods. With our approach, a fast and error-controlled exploration of chemical reaction networks is now in reach

Speaker 2 – Kamil Ciosek Title – Reinforcement Learning with Continuous Actions: Fourier Policy Gradients and Expected Policy Gradients Abstract – The talk motivates the use of RL for continuous control tasks and gives a brief overview of actor-critic methods applied in this setting. I describe recent advances in the field and discuss how these methods are derived. I discuss the advantages and disadvantages of estimators used in policy gradients and a framework for deriving them, known as Fourier Policy Gradients. I also discuss approaches to exploration used in this setting, particularly a scheme arising from the Expected Policy Gradients framework that uses the curvature of the critic to drive exploration. Finally, I give an example of how RL with continuous actions is used at Microsoft Research