This paper describes a theory of intelligent systems and its reduction to engineering practice. The theory is based on a
broader theory of computation wherein information and control are defined within the subjective frame of a system. At
its most primitive level, the theory describes what it computationally means to both ask and answer questions which, like
traditional logic, are also Boolean. The logic of questions describes the subjective rules of computation that are
objective in the sense that all the described systems operate according to its principles. Therefore, all systems are
autonomous by construct. These systems include thermodynamic, communication, and intelligent systems. Although
interesting, the important practical consequence is that the engineering framework for intelligent systems can borrow
efficient constructs and methodologies from both thermodynamics and information theory. Thermodynamics provides
the Carnot cycle which describes intelligence dynamics when operating in the refrigeration mode. It also provides the
principle of maximum entropy. Information theory has recently provided the important concept of dual-matching useful
for the design of efficient intelligent systems. The reverse engineered model of computation by pyramidal neurons
agrees well with biology and offers a simple and powerful exemplar of basic engineering concepts.
Sensor networks are fast-emerging as a powerful technology for many distributed surveillance and monitoring applications.
The crux of these applications lies in being able to compute various global characteristics, like contour, trajectory,
direction, velocity etc., of mobile phenomena. In this paper, we present algorithms based on in-network aggregation
to efficiently compute these global characteristics of phenomena by representing them as functions amenable
to in-network processing, at a very low communication cost. We present two versions of these algorithms; one for
post-event querying and the other for real-time warnings and include simulation results for all our algorithms.
In this paper, we present new concepts for path planning and cooperative target assignment for unmanned vehicles in
uncertain environments. The approach is inspired by possible cognitive computing architectures used in mammalian
brains for navigation tasks. A dynamic map based on sensor input and a diffusion equation is used to generate paths
around obstacles to targets. Targets are assigned to each UAV based on simple priority logic or dynamic competition.
Simulation results for a simple scenario with two UAVs and multiple targets are presented.
In previous work by the author, effective persistent and pervasive sensing for recognition and tracking of battlefield
targets were seen to be achieved, using intelligent algorithms implemented by distributed mobile agents over a
composite system of unmanned aerial vehicles (UAVs) for persistence and a wireless network of unattended ground
sensors for pervasive coverage of the mission environment. While simulated performance results for the supervised
algorithms of the composite system are shown to provide satisfactory target recognition over relatively brief periods of
system operation, this performance can degrade by as much as 50% as target dynamics in the environment evolve
beyond the period of system operation in which the training data are representative.
To overcome this limitation, this paper applies the distributed approach using mobile agents to the network of
ground-based wireless sensors alone, without the UAV subsystem, to provide persistent as well as pervasive sensing for
target recognition and tracking. The supervised algorithms used in the earlier work are supplanted by unsupervised
routines, including competitive-learning neural networks (CLNNs) and new versions of support vector machines
(SVMs) for characterization of an unknown target environment. To capture the same physical phenomena from
battlefield targets as the composite system, the suite of ground-based sensors can be expanded to include imaging and
video capabilities. The spatial density of deployed sensor nodes is increased to allow more precise ground-based
location and tracking of detected targets by active nodes.
The "swarm" mobile agents enabling WSN intelligence are organized in a three processing stages: detection,
recognition and sustained tracking of ground targets. Features formed from the compressed sensor data are down-selected
according to an information-theoretic algorithm that reduces redundancy within the feature set, reducing the
dimension of samples used in the target recognition and tracking routines. Target tracking is based on simplified
versions of Kalman filtration. Accuracy of recognition and tracking of implemented versions of the proposed suite of
unsupervised algorithms is somewhat degraded from the ideal. Target recognition and tracking by supervised routines
and by unsupervised SVM and CLNN routines in the ground-based WSN is evaluated in simulations using published
system values and sensor data from vehicular targets in ground-surveillance scenarios. Results are compared with
previously published performance for the system of the ground-based sensor network (GSN) and UAV swarm.
We have formulated a series of position-adaptive sensor concepts for explosive detection applications using swarms of
micro-UAV's. These concepts are a generalization of position-adaptive radar concepts developed for challenging
conditions such as urban environments. For radar applications, this concept is developed with platforms within a
UAV swarm that spatially-adapt to signal leakage points on the perimeter of complex clutter environments to collect
information on embedded objects-of-interest.
The concept is generalized for additional sensors applications by, for example, considering a wooden cart that
contains explosives. We can formulate system-of-systems concepts for a swarm of micro-UAV's in an effort to detect
whether or not a given cart contains explosives. Under this new concept, some of the members of the UAV swarm can
serve as position-adaptive "transmitters" by blowing air over the cart and some of the members of the UAV swarm can
serve as position-adaptive "receivers" that are equipped with chem./bio sensors that function as "electronic noses". The
final objective can be defined as improving the particle count for the explosives in the air that surrounds a cart via
development of intelligent position-adaptive control algorithms in order to improve the detection and false-alarm
statistics. We report on recent simulation results with regard to designing optimal sensor placement for explosive or
other chemical agent detection. This type of information enables the development of intelligent control algorithms for
UAV swarm applications and is intended for the design of future system-of-systems with adaptive intelligence for
advanced surveillance of unknown regions. Results are reported as part of a parametric investigation where it is found
that the probability of contaminant detection depends on the air flow that carries contaminant particles, geometry of the
surrounding space, leakage areas, and other factors. We present a concept of position-adaptive detection (i.e. based on
the example in the previous paragraph) consisting of position-adaptive fluid actuators (fans) and position-adaptive
sensors. Based on these results, a preliminary analysis of sensor requirements for these fluid actuators and sensors is
presented for small-UAVs in a field-enabled explosive detection environment. The computational fluid dynamics (CFD)
simulation software Fluent is used to simulate the air flow in the corridor model containing a box with explosive
particles. It is found that such flow is turbulent with Reynolds number greater than 106. Simulation methods and results
are presented which show particle velocity and concentration distribution throughout the closed box. The results indicate
that the CFD-based method can be used for other sensor placement and deployment optimization problems. These
techniques and results can be applied towards the development of future system-of-system UAV swarms for defense,
homeland defense, and security applications.
In this paper the authors attempt to address the challenges of establishing situational awareness and subsequent decision
making regarding aircraft incidents. Using Smart Nodes and its state space model it is possible to create a live threat
profile and facilitate decision making in a net-centric environment. Smart Nodes are used to monitor parts of the
environment and by combining the sensors and information fusion, create a complex and useful picture of what is and
might occur. This information needs to be quickly and efficiently disseminated to operators. The authors present a novel
method of accomplishing this specifically designed for the air transportation industry.
In this paper, activity recognition is performed based on silhouettes of the human figure obtained by background
subtraction and characterized by the shape context, a log-polar histogram derived from boundary points. In the first
approach each video frame is tagged by the activity corresponding to the closest matches between the query and known
shapes. In the second method, the shape context dimensionality is reduced by principal components analysis, and a
neural network is used for activity classification of individual frames. The overall decision for an entire video sequence
is based on majority vote. Classification of individual frames ranged between 70-90% and overall classification of video
sequences was very accurate.
Data acquisition with multiple sensors requires accurate registration in both time and space for effective data
fusion. This paper presents a system that permits synchronization for GPS and video, but it can be expanded
to include other sensors (i.e. infrared, SAR, etc). We begin with a discussion on the using a pulse-per-second
signal for synchronization. We then describe the workings of the Global Positioning System. We compare
dfferent autopilots, cameras, frame captures, processors, operating systems, and data storages that can be
used for the system, and provide our hypothesis for device selection. We introduce the overall idea of video
compression, brieify summarize the different methods, explain our choice of MPEG-2, describe the metadata
format, compare the choices of encoders, and explain the MPEG transport stream. We present the benefits of
using certain transmission frequencies and the legal restrictions, and give our frequency choices based on available
transmitters and receivers. We finish with a summary of the entire system.
In the distributed operations of route discovery and maintenance, strong interaction occurs across mobile ad hoc
network (MANET) protocol layers. Quality of service (QoS) requirements of multimedia service classes must be
satisfied by the cross-layer protocol, along with minimization of the distributed power consumption at nodes and along
routes to battery-limited energy constraints. In previous work by the author, cross-layer interactions in the MANET
protocol are modeled in terms of a set of concatenated design parameters and associated resource levels by multivariate
point processes (MVPPs). Determination of the "best" cross-layer design is carried out using the optimal control of
martingale representations of the MVPPs.
In contrast to the competitive interaction among nodes in a MANET for multimedia services using limited resources,
the interaction among the nodes of a wireless sensor network (WSN) is distributed and collaborative, based on the
processing of data from a variety of sensors at nodes to satisfy common mission objectives. Sensor data originates at the
nodes at the periphery of the WSN, is successively transported to other nodes for aggregation based on information-theoretic
measures of correlation and ultimately sent as information to one or more destination (decision) nodes. The
"multimedia services" in the MANET model are replaced by multiple types of sensors, e.g., audio, seismic, imaging,
thermal, etc., at the nodes; the QoS metrics associated with MANETs become those associated with the quality of fused
information flow, i.e., throughput, delay, packet error rate, data correlation, etc. Significantly, the essential analytical
approach to MANET cross-layer optimization, now based on the MVPPs for discrete random events occurring in the
WSN, can be applied to develop the stochastic characteristics and optimality conditions for cross-layer designs of sensor
network protocols. Functional dependencies of WSN performance metrics are described in terms of the concatenated
protocol parameters. New source-to-destination routes are sought that optimize cross-layer interdependencies to achieve
the "best available" performance in the WSN. The protocol design, modified from a known reactive protocol, adapts the
achievable performance to the transient network conditions and resource levels. Control of network behavior is realized
through the conditional rates of the MVPPs. Optimal cross-layer protocol parameters are determined by stochastic
dynamic programming conditions derived from models of transient packetized sensor data flows. Moreover, the
defining conditions for WSN configurations, grouping sensor nodes into clusters and establishing data aggregation at
processing nodes within those clusters, lead to computationally tractable solutions to the stochastic differential
equations that describe network dynamics. Closed-form solution characteristics provide an alternative to the "directed
diffusion" methods for resource-efficient WSN protocols published previously by other researchers. Performance
verification of the resulting cross-layer designs is found by embedding the optimality conditions for the protocols in
actual WSN scenarios replicated in a wireless network simulation environment. Performance tradeoffs among protocol
parameters remain for a sequel to the paper.
Sensor networks have uses ranging from personal use in homes to large-scale military applications. The ability of a sensor network is only as good as the intelligence used to control and integrate the sensors into a single entity. In this paper, we look at the marriage of agent organizations with sensor networks to create a sensor organization.
When compared to biological experiments, using computational protein models can save time and effort in identifying
native conformations of proteins. Nonetheless, given the sheer size of the conformation space, identifying the native
conformation remains a computationally hard problem - even in simplified models such as hydrophobic-hydrophilic
(HP) models. Distributed systems have become the focus of protein folding, providing high performance computing
power to accommodate the conformation space. To use a distributed system efficiently (with limited resources), an
appropriate strategy should be designed accordingly. Communication incurs overhead but can provide useful
information in distributed systems through careful consideration. Our study focuses on understanding the behavior of
distributed systems and developing an efficient communication strategy to save computational effort in order to obtain
good solutions. In this paper, we propose a distributed caching strategy, which reuses partial results of computations and
transmits the cached and reusable information among neighboring inter-connected processors. In order to validate this
idea in a practical setting, we present algorithms to retrieve and restore the cached information and apply them to 2D
triangular HP lattice models through coarse-grained parallel genetic algorithms (CPGAs). Our experimental results
demonstrate the time savings as well as the limits in caching improvements for our distributed caching strategy.
Advances in the areas of robotics have greatly increased the complexity and number of problems that groups
of robots are able to solve. This work deals with the use of homogeneous and autonomous robots dynamically
forming teams to solve a multi-threat containment problem. The multi-threat containment problem has the
robot teams surround a number of threats that may occur randomly. Approaches with and without utilizing
wireless communication are proposed and analyzed with a focus on the effects of using wireless. Simulation
results show the benefit of the proposed integrated algorithm and its performance in different scenarios.
Intelligent computing has various methods where a structure is set up and activated from outside. The separation of
structure-building and activity into different modes of operation makes the more complex problems unreachable. This
paper discusses a method of operation for intelligent computing where a structure is internally active, dynamically
extensible by the user, and also modifiable by itself, with new elements of structure immediately participating in activity.
Three themes are explored - directed versus undirected structure, a method of self-extension and the operation of free
structure. Existential control and inheritance are shown to be computable within the structure and the need to model
relations on relations for complex applications is discussed.
Appreciating terrain is a key to success in both symmetric and asymmetric forms of warfare. Training to enable Soldiers
to master this vital skill has traditionally required their translocation to a selected number of areas, each affording a
desired set of topographical features, albeit with limited breadth of variety. As a result, the use of such methods has
proved to be costly and time consuming. To counter this, new computer-aided training applications permit users to
rapidly generate and complete training exercises in geo-specific open and urban environments rendered by high-fidelity
image generation engines. The latter method is not only cost-efficient, but allows any given exercise and its conditions to
be duplicated or systematically varied over time. However, even such computer-aided applications have shortcomings.
One of the principal ones is that they usually require all training exercises to be painstakingly constructed by a subject
matter expert. Furthermore, exercise difficulty is usually subjectively assessed and frequently ignored thereafter. As a
result, such applications lack the ability to grow and adapt to the skill level and learning curve of each trainee. In this
paper, we present a heuristic that automatically constructs exercises for identifying key terrain. Each exercise is created
and administered in a unique iteration, with its level of difficulty tailored to the trainee's ability based on the correctness
of that trainee's responses in prior iterations.
A key to mastering asymmetric warfare is the acquisition of accurate intelligence on adversaries and their assets in urban
and open battlefields. To achieve this, one needs adequate numbers of tactical sensors placed in locations to optimize
coverage, where optimality is realized by covering a given area of interest with the least number of sensors, or covering
the largest possible subsection of an area of interest with a fixed set of sensors. Unfortunately, neither problem admits a
polynomial time algorithm as a solution, and therefore, the placement of such sensors must utilize intelligent heuristics
instead. In this paper, we present a scheme implemented on parallel SIMD processing architectures to yield significantly
faster results, and that is highly scalable with respect to dynamic changes in the area of interest. Furthermore, the
solution to the first problem immediately translates to serve as a solution to the latter if and when any sensors are
Persistent sensing by Unmanned Airborne Vehicles (UAVs) has brought up challenging issues including multi-scale
analysis, multi-modal sensor fusion, and scene localization. As for the first issue, the multi-scale and
multi-resolution issues occur when a mobile sensor changes altitudes or two different sensors with the same
camera provide any redundant images from different altitudes. To overcome these issues, we first focus on
collecting invariant feature data from the multi-resolution representation of a high resolution image. Recently, an
information-theoretic matching criterion has been developed for robust data registration without any knowledge
of feature correspondence. This criterion is used as an intelligent computing algorithm of choosing a good scale-representation
that helps to find an unknown scaling factor between two different and redundant measurements.
As for the second issue of multi-modal sensor fusion, we observe that Electro-optical (EO) and Infrared (IR)
images in the DARPA VIVID database have an inherent scaling-difference, even though the different modalities
come from the two fixed EO and IR sensors attached on the same mobile sensor. Here we provide a new
experimental result of multi-modal data fusion that successfully combines complementary information via the
process of data refinement. The recovered transformation reveals one of the fundamental characteristics of the
two different modalities. The last issue of scene localization is required for identifying the scene visited before.
In this paper, we demonstrate the trajectory of the mobile sensor based only on the extracted transformations
(not relying on any telemetric data of the mobile sensor which is not available persistently) by projecting the
center locations of image measurements onto the two dimensional reference coordinate.
Image registration is an important fundamental process, which has many useful applications to Tracking, Automatic
Target Recognition (ATR), and Sensor Fusion to name a few. Recent publications illustrate the Air
Force's desire to exploit the benefits of a layered (altitude) sensing environment. In such an environment, data
from different sensors, of different modalities, at different elevations of the same scene are fused to gain a better
understanding of the operational environment. This research, sponsored by the Air Force Research Lab (AFRL),
builds on classical registration techniques to explore novel registration algorithms applied to data under the new
layered sensing environment. Our main focus, herein, is to register large-scale aerial Electro Optical (EO) images
collected from cameras at different altitudes. Particularly we propose a method to jointly segment and register
the same object in two layered images. A multiphase, region-based active contour method, together with an
adapted joint segmentation-registration technique, are combined to provide a way of registering layered sensing
images. This method provides a Level 0 mechanism in data fusion hierarchy to preprocess (segment) and align
data for future fusion stages. In this paper, theories of multi-view geometry is reviewed for understanding the
possible form of registration, our proposed method is illustrated and substantiating examples and results are provided as well.
An intelligent medical systems based on a radial basis neural network is applied to the automatic classification of suspicious lesions in breast MRI and compared with two standard mammographic reading methods. Such systems represent an important component of future sophisticated computer-aided diagnosis systems and enable the extraction of spatial and temporal features of dynamic MRI data stemming from patients with confirmed lesion diagnosis. Intelligent medical systems combining both kinetics and lesions' morphology are expected to have substantial implications in healthcare politics by contributing to the diagnosis of indeterminate breast lesions by non-invasive imaging.
In industrial process control, many processes or plants are already stable. Thus, the desired process transient behavior
and steady state error are the design constraints in these cases. Two common control techniques used in process control
are internal model control (IMC) or Proportional Integral Derivative (PID) control. IMC can only be used on already
stable or stabilized plants or processes due to its structure. Many plants or processes though cannot be completely
identified or are modeled using reduced order linear models. This can lead to modeling errors. On the other hand, neural
networks can be used to identify nonlinear processes or functions. In this research, neural networks are used for
intelligent/adaptive system identification of the plant to be utilized in the internal model control. This adaptive neural
network IMC structure is simulated to control a simplified process model. The efficacy of the neural network IMC
method is compared to classic PID control.
An increase in demand for computing power in academia has necessitated the need for high performance machines.
Computing power of a single processor has been steadily increasing, but lags behind the demand for fast simulations.
Since a single processor has hard limits to its performance, a cluster of computers can have the ability to multiply the
performance of a single computer with the proper software. Cluster computing has therefore become a much sought after
technology. Typical desktop computers could be used for cluster computing, but are not intended for constant full speed
operation and take up more space than rack mount servers. Specialty computers that are designed to be used in clusters
meet high availability and space requirements, but can be costly. A market segment exists where custom built desktop
computers can be arranged in a rack mount situation, gaining the space saving of traditional rack mount computers while
remaining cost effective. To explore these possibilities, an experiment was performed to develop a computing cluster
using desktop components for the purpose of decreasing computation time of advanced simulations. This study indicates
that small-scale cluster can be built from off-the-shelf components which multiplies the performance of a single desktop
machine, while minimizing occupied space and still remaining cost effective.