Computer Science Seminar Presentation

6 pm, March 25th, 2015, Del Norte 2530

Identification of Location Cues for Robot Navigation with a Laser Scanner and Python Data Mining Tools

Prof. AJ Bieszczad

Computer Science, Channel Islands
California State University

Abstract: In this talk, I report on our attempts to build a cue identifier for mobile robot navigation using machine learning techniques. We use simulated 2D LIDAR laser scanner data to identify techniques that are most promising, and then apply the model to data in a larger dimension space. We apply various levels of noise to the ideal scanner images of environmental cues to accommodate laser inaccuracy. Additionally, we consider various points of view. Our results indicate that good models can be built with both back-propagation neural network applying Broyden-Fletcher-Goldfarb-Shannon (BFGS) optimization and regularization, and with Support Vector Machines (SVMs) assuming that data shaping took place with a normalization followed by a Principal Component Analysis (PCA). We show that expanding data dimension thirty-fold has acceptable impact of the training time of SVM models. We will use iPython Notebook with the scientific environment based on NumPy, SciPy, Matplotlib, Scikit Learn, and Neurolab, to illustrate the experiments.

AiderLaser Scan