Research & Development

Evidential Reasoning Network (ERN)

In the development of complex decision support systems, 21CSI has addressed the need for combining the information provided by drastically different information sources (various sensors, intelligence reports, data mining systems, etc.) to develop a set of recommended courses of action for a given situation. Evidential reasoning (ER) provides a mechanism for reasoning under uncertainty and drawing conclusions from multiple sources of evidence and opinion. Since the introduction of the Dempster-Shafer Theory of Evidence, new ER methods have been and are continuing to be developed, including Subjective Logic. An evidential reasoning framework was needed to ensure that evidential reasoning expressions are coherent, consistent, and computationally tractable. 21CSI’s Evidential Reasoning Network (ERN) architecture is a novel structure that addresses these needs. ERN can automatically derive supporting evidence from multiple sources including opinions from trusted sources, previous judgments, and ontological knowledge. ERN can also be used as a vehicle for knowledge discovery and as a tool for assessing what a system knows. ERN uses a belief algebra structure for providing a mathematically rigorous representation and manipulation of uncertainty within the evidential reasoning network. The belief algebra structure is capable of using probabilistic belief mass assignments through the use of belief frames. The two prime belief algebra operators required are consensus and discount. The consensus operator allows the combination of belief values through the network amongst various opinion generating authorities, such as human subject matter experts or software agents that perform some sort of data analysis and reasoning. The discount operator is used in the structure of trust chains and in the propagation of belief values based on saliency and trust. Trust chains are a significant advantage 21CSI’s ERN provides over most existing data fusion systems. There is a fundamental difference between the existential value of a belief statement and the belief in the authority making the statement. Data fusion architectures ignoring this dichotomy can create needlessly complex networks and provide poor support for pedigree management and failsafe mechanisms. In the ERN architecture, authorities can be agents or humans, and there is no real architectural distinction in how an authority’s belief statement is handled. Patent pending.

Adaptive Resonance Theory Engine (ART-E)

21st Century Systems, Inc., along with researchers from the Missouri University of Science and Technology, have been developing the ART-E technology to handle the data clustering needs of a diverse problem set that works well on small datasets, yet scales well to large datasets. Adaptive resonance theory (ART) was developed by Carpenter and Grossberg to learn arbitrary input patterns in a stable, fast, and self-organizing way, thus overcoming the effect of learning instability. It is a learning theory hypothesizing that resonance in neural circuits can trigger fast learning. We initiated development of ART-E to address the challenge of identifying Anthrax dispersal models from sparse data sets. Following this initial success, we modified the learning scheme to handle the far more challenging domain of target recognition under bomb damage. In this case, we implemented a complement coding scheme to address the condition where large pieces of the structure are missing from the scene after a strike. The resulting engine was able to identify 80% of targets up to 48% total damage. Our research most recently has applied the ART-E technology to the intelligence, surveillance, and reconnaissance (ISR) domain, where it was used as data characterizer. We are able to characterize the data from a set of features (ranging from semantic textual identifiers to physical data phenomena). Once characterized, 21CSI uses this information in our decision support software applications to provide situational awareness and decision support.

VisionAgent®

21CSI’s VisionAgent® is a network-based software agent that monitors a live video stream and provides XML-based alerts to subscribing systems. VA has been adapted for the land-based surveillance needs of our force protection programs, including classifiers used to detect people and vehicles. Derived from SBIR technologies, VisionAgent performs video analysis on a live video stream (fixed/mobile camera, UAV sensor, etc.). The current algorithm chain includes processes for image stabilization, change detection, characterizers, Kalman Filtering, and much more. VisionAgent uses techniques that are fast enough so that real-time analysis can be done. VisionAgent® is used as a research platform to develop entirely new algorithms and as a commercial platform in existing products. It is made up of a set of building block modules. These building block modules are strung together to form a VisionAgent chain for a particular application. The basic building blocks that are being developed fall into one of these categories: Preprocessing, Detection, Classification, Tracking, and Decision/Action. The VisionAgent tools are divided into two types: Enhanced Video and Smart Video. Enhanced video is a technology that processes and improves the quality of video by enhancing the signal while dampening the noise. Video enhancement processes do not add artificial content, therefore maintaining the integrity of the original information. Smart video is the technology of processing the video to extract information. For instance, it is used to find objects, patterns, and make decisions about the video. It can autonomously attempt to determine what is in the video, taking the human out of the loop, or aid humans to be alerted of objects or events to allow surveillance decisions to be made faster.

FUsion Through Uncertainty REduction (FUTURE)

FUTURE is derived from a Phase I SBIR effort. In this Phase I, we successfully demonstrated the effectiveness of fusion algorithms that analyzed passive signal properties. Utilizing internal research and development funds, we extended that initial research and developed the FUTURE fusion algorithm. The FUTURE algorithm is a credible and possibly superior technique when compared to contemporary statistical search/optimization methods. FUTURE has an advantage over existing methods because it recognizes and uses to its advantage the fact that not all sensor detection measurements are created equal. This is embodied within the mathematics of FUTURE as an ambiguity measure where each measurement (either passive or active, monotonic, bistatic, or multistatic) is measured with an ambiguity coefficient and the measurements are sorted in ascending order of that coefficient. This puts the “easiest to explain” measurement at the top of the sorted list. FUTURE iterates through this list fusing the data, much like a Sudoku puzzle is solved, the obvious answers determined first leading to a clearer picture for the more complex solutions.

PinPST – Paths in Perforated Space Time

PinPST solves path finding problems by expressing the search space as a space-time volume. Since any principal-object (e.g., ship, robot, etc.) moving in space sweeps out a trajectory-volume in space-time, then any path finding problem is equivalent to finding the trajectory-volume which best satisfies all the problem constraints. The space-time volume through which the object may move is limited by the maximum speed of the object. Thus the principal-object’s trajectory is confined to the interior of a space-time speed-cone (STSC) on which a directed trellis graph is constructed. Furthermore any known or hypothesized avoidance-object also describes a trajectory in space-time whose intersection with the STSC is a region forbidden to the principal-object. Thus removing all avoidance trajectories leaves a perforated STSC (or a perforated trellis graph), which is a feasible path region. Perforating the STSC with avoidance trajectories has the affect of pruning the trellis graph. In fact the more objects there are to avoid the more the trellis is pruned, with the implication being that the more complex the environment, the faster the algorithm executes. Further graph pruning using principal-object operating characteristics (e.g., maximum speed, turning radius, etc.) is performed. Nodes and edges can have problem dependent weights (e.g., probability of finding a threat object, cost to move in that direction, etc). The solution path is then determined using a hybrid depth-first and D* (i.e., dynamic A*) search on the reduced and weighted trellis.