Team Flowers

Members
Overall Objectives
Scientific Foundations
Application Domains
Software
New Results
Contracts and Grants with Industry
Other Grants and Activities
Dissemination
Bibliography

Section: New Results

ILO-GMR: Incremental Local Online Gaussian Mixture Regression

Participants : Ming Li, Pierre-Yves Oudeyer, Adrien Baranès.

Many robot learning frameworks involve the use of regression algorithms. In such frameworks, the desirable properties of these algorithms are: 1) they should be able to work in high-dimensions; 2) They should be fast to train with millions of points; 2) Training should be incremental; 3) prediction should be fast; 4) They should be easy to manually tune when shifting from one problem to another. A number of techniques have become popular recently in robotics, such as Gaussian Processes, Locally Weighted Projection Regression (LWPR) and Gaussian Mixture Regression (GMR). But Gaussian Processes are often slow to train and not incremental, while LWPR is very difficult to tune because of its many parameters and GMR has no efficient incremental versions and need to be retrained globally when new training data is provided with a different input distribution. We have elaborated an incremental online local version of GMR, based on the only computation of local GMR with few components, which we proved to be as accurate as the other methods, robust to changes in the distribution of the training data, as well as very easy to tune. An extensive article describing the algorithm and its performances is in preparation.


previous
next

Logo Inria