Data-Efficient Scalable Reinforcement Learning for Practical Robotic Environments

Informations projet

DESlRE

N° de convention de subvention: 798321

DOI

10.3030/798321

Projet clôturé

Date de signature de la CE 19 Mars 2018

Date de début 1 Avril 2018

Date de fin 31 Mars 2020

Financé au titre de

• EXCELLENT SCIENCE - Marie Skłodowska-Curie Actions

Coût total

€ 159 460,80

Contribution de l’UE

€ 159 460,80

Coordonné par

MAX-PLANCK-GESELLSCHAFT ZUR FORDERUNG DER WISSENSCHAFTEN EV
Germany

Periodic Reporting for period 1 - DESlRE (Data-Efficient Scalable Reinforcement Learning for Practical Robotic Environments)

Période du rapport: 2018-04-01 au 2020-03-31

Résumé du contexte et des objectifs généraux du projet

The initial aim of the project was to develop algorithms suitable for challenging control tasks. Current algorithms that perform well in simulation typically transfers suboptimally to real test or robots. For example, one of the hot topics in research is termed- sim-to-real transfer, which aims to transfer amazing feats that algorithms can achieve in computer simulations to test time performance. If we can understand how to perform control in a highly stochastic environment, many problems in social decision-making can be solved. The overall objectives are to advance our understanding of the difficulty of such an application of algorithms and developing new ones.

Travail effectué depuis le début du projet jusqu’à la fin de la période considérée dans le rapport et principaux résultats atteints jusqu’à présent

The fellow has developed various new optimization and control algorithms that are designed to handle the hybrid and stochastic environment. Our theory connects to the deep theory of robust optimization and robust control, as well as the robustness of machine learning algorithms such as kernel method. One of the most significant research outcomes was the newly proposed framework of kernel distributionally robust optimization algorithm. This elegant framework is a combination of principled robust optimization theory and kernel machine learning.

Progrès au-delà de l’état des connaissances et impact potentiel prévu (y compris l’impact socio-économique et les conséquences sociétales plus larges du projet jusqu’à présent)

Fellow's work presents new insights combining the theory of robust convex optimization and RKHS. It shows the theory of kernel methods can be used to make robust decisions for general decision-making problems. The work adds an interesting piece to both DRO and kernel method literature.

From a practical perspective, we have proposed easy-to-implement algorithms. As we discussed in recent works, one strength of our methods is its wide applicability. Many of today's learning tasks suffer from manifestations of distributional ambiguity. We believe practitioners from industry and business that wish to gain robustness in their learning or decision-making tasks can apply our kernel distributionally robust optimization algorithms.

We aim to design optimization and control algorithm that can hedge against distribution shift