Autonomous decision making system using Genetic Programming and Neural Networks
One of the greatest challenges in the detection/classification technologies is to develop a decision making algorithm, because signatures from targets and clutters are often very similar and requires extremely sophisticated processes to discriminate the target from clutter signatures. Autonomous decision making algorithms have not reached the stage of practical use due to complexity in interpreting both quantitative and qualitative features of targets and large amount of data. Only limited detection algorithms were demonstrated by human expertise. This process is often time consuming and possesses the potential of error associated with human factors. It is essential to develop automatic decision making algorithm for fast, robust, and objective detection/classification task.

Currently, our group is investigating several artificial intelligent (AI) techniques to automate the classification algorithms. In particular, Neural Network (NN), historically well known for capability of pattern recognition and classification, and the Genetic programming (GP), relatively new method of the AI techniques, have been examined to realize the decision making process of autonomous classification algorithm. The NN and the GP techniques have been compared using a set of test images (see Fig. 1a) from the classification point of view, and advantages and disadvantages of each technique were evaluated. The GP is found (see Fig. 1b) to have out performed NN methods in highly noisy environments and by large margin for cases when incomplete target training was used. It is believed that GP methods will play a significant role in autonomous classification algorithms. Thus, we propose to continue this research effort particularly when arming the GP algorithm with realistic features and trends based on field data.

Deliverables & milestones:

  • Development of the test classification algorithms using NN and GP, respectively.
  • Evaluate classification performance of the NN and GP.
  • Development of decision making rules for target classification based on actual field data.
  • Apply GP and NN to realize the autonomous classification algorithm.
  • Evaluate performance of the autonomous classification algorithm by blind test.
Fig. 1. (a) Example of test images used for character recognition with additive white noise ranging from 0% to 50% . (b) Character recognition performances of 4 different structures of NN and GP. Six alphabets with additive white noise were used as an examinee. Plot shows that, for all NNs and GPs, the error rate increases as the noise level increases. GP provides the best recognition performance.

HCAC | University of Hawaii | College of Engineering |1680 East West Rd.| Post 201 | Honolulu HI 96822 | PH:(808) 956-3434

 

UH Manoa College of Engineering HCAC homepage University of Hawaii at Manoa