Battlespace Situation Assessment via Clustering and Case-Base Reasoning by Carl G. Looney* and Lily R. Liang Computer Science Department/171 University of Nevada, Reno, NV 89557 <looney,liang>@cs.unr.edu Abstract. We cluster surface target feature vectors by position in a certain area of the battlespace and make inventories of the resulting clusters by type and count. The feature vectors come from the target tracks. Our new centralized-mean clustering method is robust. Next, we apply case-based reasoning to infer the enemy unit types and their posture for situation awareness. We then employ a weighted retrieval process to match the new cluster inventories to cases whose solutions provide unit types. 1. Introduction Situation awareness is a model of the the locations, types, counts, readiness and activity levels of enemy resources, along with extraneous information on the terrain, weather, access routes, etc. Situation assessment (SA) is the ongoing process of inferring this relevant information about forces of concern in a military situation. Commanders need visual models for spatial reasoning and decision making. They need situation awareness [5] of their battlespaces to be able to make effective command decisions [9]. History shows that it is a key to victory and that the lack of it often leads to disaster [4]. A visual model of the battlespace allows the commander and his forces to operate according to a common perception. Perception is a process [3] of maintaining a model of an environment by combining observed and stored data into a coherent description. The summarized data comprise a state of the battlespace environment that consists of its objects and their interactivity. It is updated based on the previous state data, newly observed data (and possibly other incoming information) and models of interactivity. The data for SA comes from the tracks of targets that are established by readings from sensors such as radar, infrared, laser, optic and other sensors (electronic intelligence and communications intelligence). The sensor platforms may be ground, sea, air or space based. The track data fields contain time, position, target type, and sometimes subtype. A track may contain not only the latest state, but a short sequence of previous states. ______________________________________________ * Supported by ARO Grant DAAD19-99-1-0089 The ground (or sea) situation picture is part of the common operational picture of the greater batttlespace, maintained by the friendly forces using data fusion [9]. The situation should be interpretable as patterns that are recognizable from commanders’ experience and intuition. Situation awareness is an enabler for gaining operational advantage, but a profusion of data from sensors, communications reports and databases can be counterproductive and contribute to confusion (the fog of war becomes the glare of war). This delays the time critical decision making required to operate inside the response cycle of the enemy. 2. Clustering the Targets The feature vector for each surface target contains the information shown in Table 1. The (x,y,z)-position components are given in real (floating point) numbers to designate kilometers and the class beliefs are reals between 0.0 and 1.0, but everything else is integer valued. The feature vectors for a battlespace are to be updated from tracks at certain times to yield a current set of Q target feature vectors {x (q) : q = 1,...,Q} in various surface areas for SA processing. The clustering of these vectors is done only with respect to position features (x,y) or (x,y,z) in a given area of 2x2 kilometers (or other size where a nonnegligible number of ground targets are located). Our robust clustering algorithm forms clusters and determines their centroids in a manner that is not influenced by outliers and noisy vectors [8]. Our vector averaging method is similar in effect to the alpha-trimmed mean [2] used for real values. The clustering of the feature vectors, denoted by {x (q) : q = 1,...,Q}, is started with a relatively large number K of uniformly randomly drawn seed vectors {z (k) : k = 1,...,K}. We thin this large set to obtain a smaller set of uniformly distributed seeds for the initial centroids of clusters. Starting with the first seed, any seeds closer to it than the threshold J are eliminated, K is decremented and the next available seed is checked the same way, and so forth. This prevents bad seeds from causing a bad clustering [8].