Warning:
This page is provided for historical and archival purposes only. While the seminar dates are correct, we offer no guarantee of informational accuracy or link validity. Contact information for the speakers, hosts and seminar committee are certainly out of date.

RI Seminar -- Andrew W. Moore

• WHEN: Friday, April 29, 1994; 3:30 pm - 5:00 pm Refreshments will be served starting at 3:15 pm
• WHERE: ADAMSON WING Auditorium in Baker Hall
• SPEAKER: Andrew W. Moore

          The Robotics Institute
          Carnegie Mellon University
         

• TITLE: Learning control with kernel-based function approximators and intense cross-validation

ABSTRACT

A learning controller can benefit considerably from explicitly remembering every experience in its lifetime. In this talk I will discuss how locally weighted regression and its variants can be used for controlling robots and other complex systems. I will then describe how models of forward and inverse dynamics learned together can provide a robust training regime.

While learning control, data can sometimes be more plentiful than in typical function approximation applications. Cheaply available data can be used to further increase the autonomy of the learning controller. Intense searches can be carried out for finding parameters, regression-orders and subsets of features which minimize cross-validation error. Much of this talk will concern methods for doing this with computational efficiency, broadly based on the idea of quickly cutting off cross-validation computations which are not predicted to give useful results. Improvements are then given, including (1) the use of blocking to quickly spot near-identical models, and (2) schemata search: a new method for quickly finding families of relevant features. Experiments are presented for robot data and noisy synthetic datasets. The new algorithms speed up computation without sacrificing reliability, and in some cases are more reliable than conventional techniques.

If time permits I will also discuss a new kd-tree algorithm for accelerating the cost of locally weighted regression predictions. This algorithm permits points to be added incrementally to the tree with logarithmic cost, and is particularly appropriate for input spaces with many variables or for local weighting functions which are not very local (perhaps using a significant fraction of the stored datapoints).

Host:           Yangsheng Xu (xu@cs.cmu.edu)
Appointment:    Lalit Katragadda (lalit@cs.cmu.edu)