Improving the performance of ANN training with an unsupervised filtering method

  • Authors:

  • Sekou Remy;Chung Hyuk Park;Ayanna M. Howard

  • Affiliations:

  • Human-Automation Systems Lab, School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA;Human-Automation Systems Lab, School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA;Human-Automation Systems Lab, School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA

  • Venue:
  • IJCNN'09 Proceedings of the 2009 international joint conference on Neural Networks
  • Year:
  • 2009

Quantified Score

Hi-index 0.00

Visualization

Abstract

Learning control strategies from examples has been identified as an important capability for many robotic systems. In this work we show how the learning process can be aided by autonomously filtering the training set provided to improve key properties of the learning process. Demonstrated with data gathered for manipulation tasks, the results herein show the improved performance when autonomous filtering is applied. The filtration method, with no prior knowledge of the task was able to partition the training sets into sets almost equal to expertly labeled sets. In the case where the filter did not produce the same groupings as the expert user, the method still permitted a controller to be trained which demonstrated a success rate of 92%.