Informative Path Planning for Improving Classification Performance through Kinematic Decisions

Abstract

As an information-gathering agent, unmanned system inevitably relies on its sensing abilities to perceive the world. Such a gathered information is vital for making further significant decisions, thus often operating under the assumption that maximizing the amount of information would result in minimizing the likelihood of committing erroneous decisions. In this work, we question this assumption by carefully examining the relationship between information and the probability of making erroneous decisions, then investigate the implication of the relationship in a Unmanned Air Vehicle (UAV) path planning problem. First, we conduct a functional analysis of two performance metrics (i.e., mutual information and the probability of misclassification) with respect to the sensing abilities. The analysis suggests an effective region of sensor space that can improve the classification performance when multiple measurements are to be taken sequentially. Based on the results of the analysis, we establish sensing strategies to a UAV path planning problem where the sensor performance depends on the relative position (i.e., range and azimuth) of the UAV with respect to the object of interest. Specifically, we use two sliding mode controllers, each of which accounts for a particular sensing strategy, with a hybrid-system switching scheme. We validate our approach with numerical simulation results.