Empirical Learning Using Rule Threshold Optimization for Detection of Events in Synthetic Images

Abstract

We have developed an expert system for interpretation of passive sonar images. A key component of the system is a group of event detection rules whose conditions consist of tests against thresholds. Due to the complexity, variability and clumpiness (i.e., tendency towards highly nonuniform distribution) of the data, tuning these thresholds for good performance under all conditions is a difficult task. We have implemented a procedure for learning rule thresholds whereby the detection capability of each rule continually improves as more and more data is played through the system. The learning procedure contains the following components: 1) a windowing mechanism that adds exceptions (i.e., false alarms and missed detections) into a training database of positive and negative examples and 2) a genetic algorithm to optimize the thresholds with respect to the training database. The genetic training algorithm allows the developer to explicitly choose an operating point on the Receiver Operating Characteristic (ROC) curve of a rule. Experiments have verified 1) the superiority of this automated approach to selecting rule thresholds over manual techniques and 2) the improvement of rule performance with experience.