Search

A Bayesian network approach to self-organization and learning is introduced for use with intelligent agents. Bayesian networks, with the help of influence diagrams, are employed to create a decision-theoretic intelligent agent. Influence diagrams combine both Bayesian networks and utility theory. In this research, an intelligent agent is modeled by its belief, preference, and capabilities attributes. Each agent is assumed to have its own belief about its environment. The belief aspect of the intelligent agent is accomplished by a Bayesian network. The goal of an intelligent agent is said to be the preference of the agent and is represented with a utility function in the decision theoretic intelligent agent. Capabilities are represented with a set of possible actions of the decision-theoretic intelligent agent. Influence diagrams have utility nodes and decision nodes to handle the preference and capabilities of the decision-theoretic intelligent agent, respectively. Learning is accomplished by Bayesian networks in the decision-theoretic intelligent agent. Bayesian network learning methods are discussed intensively in this paper. Because intelligent agents will explore and learn the environment, the learning algorithm should be implemented online. None of the existent Bayesian network learning algorithms has online learning. Thus, an online Bayesian network learning method is proposed to allow the intelligent agent learn during its exploration. Self-organization of the intelligent agents is accomplished because each agent models other agents by observing their behavior. Agents have belief, not only about environment, but also about other agents. Therefore, an agent takes its decisions according to the model of the environment and the model of the other agents. Even though each agent acts independently, they take the other agents behaviors into account to make a decision. This permits the agents to organize themselves for a common task. To test the proposed intelligent agent's learning and self-organizing abilities, Windows application software is written to simulate multi-agent systems. The software, IntelliAgent, lets the user design decision-theoretic intelligent agents both manually and automatically. The software can also be used for knowledge discovery by employing Bayesian network learning a database. Additionally, we have explored a well-known herding problem to obtain sound results for our intelligent agent design. In the problem, a dog tries to herd a sheep to a certain location, i.e. a pen. The sheep tries to avoid the dog by retreating from the dog. The herding problem is simulated using the IntelliAgent software. Simulations provided good results in terms of the dog's learning ability and its ability to organize its actions according to the sheep's (other agent) behavior. In summary, a decision-theoretic approach is applied to the self-organization and learning problems in intelligent agents. Software was written to simulate the learning and self-organization abilities of the proposed agent design. A user manual for the software and the simulation results are presented. This research is supported by the Office of Naval Research with the grant number N00014-98-1-0779. Their financial support is greatly appreciated. Ph. D.

This dissertation explores the modular learning in artificial neural networks that mainly driven by the inspiration from the neurobiological basis of the human learning. The presented modularization approaches to the neural network design and learning are inspired by the engineering, complexity, psychological and neurobiological aspects. The main theme of this dissertation is to explore the organization and functioning of the brain to discover new structural and learning inspirations that can be subsequently utilized to design artificial neural network. The artificial neural networks are touted to be a neurobiologicaly inspired paradigm that emulate the functioning of the vertebrate brain. The brain is a highly structured entity with localized regions of neurons specialized in performing specific tasks. On the other hand, the mainstream monolithic feed-forward neural networks are generally unstructured black boxes which is their major performance limiting characteristic. The non explicit structure and monolithic nature of the current mainstream artificial neural networks results in lack of the capability of systematic incorporation of functional or task-specific a priori knowledge in the artificial neural network design process. The problem caused by these limitations are discussed in detail in this dissertation and remedial solutions are presented that are driven by the functioning of the brain and its structural organization. Also, this dissertation presents an in depth study of the currently available modular neural network architectures along with highlighting their shortcomings and investigates new modular artificial neural network models in order to overcome pointed out shortcomings. The resulting proposed modular neural network models have greater accuracy, generalization, comprehensible simplified neural structure, ease of training and more user confidence. These benefits are readily obvious for certain problems, depending upon availability and usage of available a priori knowledge about the problems. The modular neural network models presented in this dissertation exploit the capabilities of the principle of divide and conquer in the design and learning of the modular artificial neural networks. The strategy of divide and conquer solves a complex computational problem by dividing it into simpler sub-problems and then combining the individual solutions to the sub-problems into a solution to the original problem. The divisions of a task considered in this dissertation are the automatic decomposition of the mappings to be learned, decompositions of the artificial neural networks to minimize harmful interaction during the learning process, and explicit decomposition of the application task into sub-tasks that are learned separately. The versatility and capabilities of the new proposed modular neural networks are demonstrated by the experimental results. A comparison of the current modular neural network design techniques with the ones introduced in this dissertation, is also presented for reference. The results presented in this dissertation lay a solid foundation for design and learning of the artificial neural networks that have sound neurobiological basis that leads to superior design techniques. Areas of the future research are also presented. Ph. D.

The main goal of this research is the automatic construction of a computer model of 3D solid objects from a single range image. This research has many real world applications, including robotic environments and the inspection of industry parts. The most common methods for 3D-object extraction are based on stereo reconstruction and structured light analysis. The first approach encounters the difficulty of finding a correspondence of points between two images for the same scene, which involves intensive computations. The latter, on the other hand, has limitations and difficulties in object extraction, namely, inferring information about 3D objects from a 2D image. In addition, research in 3D-object extraction up to this point has lacked a thorough treatment of overlapped (occluded) objects. This research has resulted in a system that can extract multiple polyhedral objects from a single range image. The system consists of several parts: edge detection, segmentation, initial vertex extraction, occlusion detection, grouping faces into objects, and object representation. The problem is difficult especially when occluded objects are present. The system that has been developed separates occluded objects by combining evidence of several types. In the edge detection algorithm, noise reduction for range images is treated first by implementing a statistically robust technique based on the least median of squares. Three approaches to edge detection are presented. One that detects change in gradient orientation is a novel approach, which is implemented in the algorithm due to its superior performance, and the other two are extensions of work by other researchers. In general, the performance of these edge detection methods is considerably better than many others in the domain of range image segmentation. A hybrid approach (region-edge based) is introduced to achieve a robust solution for a single range image segmentation. The segmentation process depends on collaborating edge and region techniques where they give complementary information about the scene. Region boundaries are improved using iterative refinement. A novel approach for initial vertex extraction is presented to find the vertices of the polyhedral objects. The 3D vertex locations for the objects are obtained through an analysis of two-dimensional (2D) region shape and corner proximity, and the vertices of the polyhedra are extracted from the individual faces. There are two major approaches for dealing with occlusion. The first is an automatic identification of layers of 3D solid objects within a single range image. In this novel approach, a histogram of the distance values from a given range image is clustered into separate modes. Ideally, each mode of the histogram will be associated with one or more surfaces having approximately the same distance from the sensor. This approach works well when the objects are lying at different distances from the sensor, but when two or more objects are overlapped and lying at the same distance from the sensor, this approach has difficulty in detecting occlusion. The second approach for occlusion detection is considered the major contribution of this work. It detects occlusion of 3D solid objects from a single range image using multiple sources of evidence. This technique is based on detecting occlusion that may be present between each pair of adjacent faces associated with the estimated vertices of the 3D objects. This approach is not based on vertex and line labeling as other approaches are; it utilizes the topology and geometrical information of the 3D objects. After occlusion detection, faces are grouped into objects according to their adjacency relations and the absence or presence of occlusion between them. The initial vertex estimates are improved significantly through a global optimization procedure. Finally, models of the 3D objects are represented using the boundary representation technique that makes use of the region adjacency graph (RAG) paradigm. The experimental results of this research were obtained using real range images obtained from the CESAR lab at Oak Ridge National Laboratory. These images were obtained using a Perceptron laser range finder. These images contain single and multiple polyhedral objects, and they have a size of 512x512 pixels and a quantization of 12 bits per pixel. A quantitative evaluation of the construction algorithms is given. Part of this evaluation depends on the comparison between the results of the proposed segmentation technique and the ground truth database for these range images. The other part is to compare the results of the implemented algorithms with the results of other researchers, and it is found that the system developed here exhibits better performance in terms of the accuracy of the boundaries for the regions of the segmented images. A subjective comparison of the new edge detection methods with some traditional approaches is also provided for the set of range images. An evaluation of the new approach to occlusion detection is also presented. A recommendation for future work is to extend this system to involve images contain objects with curved surfaces. With some modifications to the multiple evidence-based approach of occlusion detection, the curved objects could be addressed. In addition, the model could be updated to include representation of the hidden surfaces for the 3D objects. This could be achieved by using multiple views for the same scene, or through assumptions such as symmetry to infer the shape of the hidden portion of the objects. Ph. D.

Visual tracking is one of the most important applications of computer vision. Several tracking systems have been developed which either focus mainly on the tracking of targets moving on a plane, or attempt to reduce the 3-dimensional tracking problem to the tracking of a set of characteristic points of the target. These approaches are seriously handicapped in complex visual situations, particularly those involving significant perspective, textures, repeating patterns, or occlusion. This dissertation describes a new approach to visual tracking for monocular and binocular image sequences, and for both passive and active cameras. The method combines Kalman-type prediction with steepest-descent search for correspondences, using 2-dimensional affine mappings between images. This approach differs significantly from many recent tracking systems, which emphasize the recovery of 3-dimensional motion and/or structure of objects in the scene. We argue that 2-dimensional area-based matching is sufficient in many situations of interest, and we present experimental results with real image sequences to illustrate the efficacy of this approach. Image matching between two images is a simple one to one mapping, if there is no occlusion. In the presence of occlusion wrong matching is inevitable. Few approaches have been developed to address this issue. This dissertation considers the effect of occlusion on tracking a moving object for both monocular and binocular image sequences. The visual tracking system described here attempts to detect occlusion based on the residual error computed by the matching method. If the residual matching error exceeds a user-defined threshold, this means that the tracked object may be occluded by another object. When occlusion is detected, tracking continues with the predicted locations based on Kalman filtering. This serves as a predictor of the target position until it reemerges from the occlusion again. Although the method uses a constant image velocity Kalman filtering, it has been shown to function reasonably well in a non-constant velocity situation. Experimental results show that tracking can be maintained during periods of substantial occlusion. The area-based approach to image matching often involves correlation-based comparisons between images, and this requires the specification of a size for the correlation windows. Accordingly, a new approach based on moment invariants was developed to select window size adaptively. This approach is based on the sudden increasing or decreasing in the first Maitra moment invariant. We applied a robust regression model to smooth the first Maitra moment invariant to make the method robust against noise. This dissertation also considers the effect of spatial quantization on several moment invariants. Of particular interest are the affine moment invariants, which have emerged, in recent years as a useful tool for image reconstruction, image registration, and recognition of deformed objects. Traditional analysis assumes moments and moment invariants for images that are defined in the continuous domain. Quantization of the image plane is necessary, because otherwise the image cannot be processed digitally. Image acquisition by a digital system imposes spatial and intensity quantization that, in turn, introduce errors into moment and invariant computations. This dissertation also derives expressions for quantization-induced error in several important cases. Although it considers spatial quantization only, this represents an important extension of work by other researchers. A mathematical theory for a visual tracking approach of a moving object is presented in this dissertation. This approach can track a moving object in an image sequence where the camera is passive, and when the camera is actively controlled. The algorithm used here is computationally cheap and suitable for real-time implementation. We implemented the proposed method on an active vision system, and carried out experiments of monocular and binocular tracking for various kinds of objects in different environments. These experiments demonstrated that very good performance using real images for fairly complicated situations. Ph. D.

Reducing the operating costs of chemical processes is very beneficial in decreasing a company's bottom line numbers. Since chemical processes are usually run in steady-state for long periods of time, saving a few dollars an hour can have significant long term effects. However, the complexity involved in most chemical processes from nonlinear dynamics makes them difficult processes to optimize. A nonlinear, open-loop unstable system, called the Tennessee Eastman Chemical Process Control Problem, is used as a test-bed problem for minimization routines. A decentralized controller is first developed that stabilizes the plant to set point changes and disturbances. Subsequently, a genetic algorithm calculates input parameters of the decentralized controller for minimum operating cost performance. Genetic algorithms use a directed search method based on the evolutionary principle of "survival of the fittest". They are powerful global optimization tools; however, they are typically computationally expensive and have long convergence times. To decrease the convergence time and avoid premature convergence to a local minimum solution, an auxiliary fuzzy logic controller was used to adapt the parameters of the genetic algorithm. The controller manipulates the input and output data through a set of linguistic IF-THEN rules to respond in a manner similar to human reasoning. The combination of a supervisory fuzzy controller and a genetic algorithm leads to near-optimum operating costs for a dynamically modeled chemical process. Master of Science

With the advent of the Global Position System (GPS), we now have the ability to determine absolute position anywhere on the globe. Although GPS systems work well in open environments with no overhead obstructions, they are subject to large unavoidable errors when the reception from some of the satellites is blocked. This occurs frequently in urban environments, such as downtown New York City. GPS systems require at least four satellites visible to maintain a good position 'fix' . Tall buildings and tunnels often block several, if not all, of the satellites. Additionally, due to Selective Availability (SA), where small amounts of error are intentionally introduced, GPS errors can typically range up to 100 ft or more. This thesis proposes several methods for improving the position estimation capabilities of a system by incorporating other sensor and data technologies, including Kalman filtered inertial navigation systems, rule-based and fuzzy-based sensor fusion techniques, and a unique map-matching algorithm. Master of Science

This thesis describes the development of a new zero-turn-radius (ZTR) differentially driven robotic vehicle hereinafter referred to as NEVEL. The primary objective of this work was to develop a device that could be used as a test-bed for continued autonomous vehicle research at Virginia Tech while meeting the entry requirements of the Annual International Unmanned Ground Robotics Competition. In developing NEVEL, consideration was given to the vehicle's mechanical and electrical design, sensing and computing systems, and navigation strategy. Each of these areas was addressed individually, but always within the context of optimal integration to produce the best overall vehicle system. A constraint that directed much of the design process was the desire to integrate industrially available and proven components rather than creating custom designed systems. This thesis also includes a review of the relevant literature as it pertains to both subsystem and overall vehicle design. NEVEL, the vehicle that was created from this research effort, is novel in several respects. It is one of the few true embodiments of a fully functioning, three-wheel, differential drive autonomous vehicle. Several previous studies have developed this concept for indoor applications, but none has resulted in a working test-bed that can be applied to an unstructured, outdoor environment. NEVEL also appears to be one of the few autonomous vehicle systems to fully incorporate a commercially available laser range finder. These features alone would make NEVEL a useful platform for continued research. In addition, however, by using common, off-the-shelf components and a personal computer platform for all computation and control, NEVEL has been created to facilitate testing of new navigation and control strategies. As testimony to the success of this design, NEVEL was recognized at the Sixth Annual International Unmanned Ground Robotics Competition as the best overall design. Master of Science

Vehicle acoustic signals have long been considered as unwanted traffic noise. In this research acoustic signals generated by each vehicle will be used to detect its presence and classify its type. Circular arrays of microphones were designed and built to detect desired signals and suppress unwanted ones. Circular arrays with multiple rings have an interesting and important property that is constant sidelobe levels. A modified genetic algorithm that can work directly with real numbers is used in the circular array design. It offers more effective ways to solve numerical problems than a standard genetic algorithm. In classifier design two main paradigms are considered: multilayer perceptrons and adaptive fuzzy logic systems. A multilayer perceptron is a network inspired by biological neural systems. Even though it is far from a biological system, it possesses the capability to solve many interesting problems in variety fields. Fuzzy logic systems, on the other hand, were inspired by human capabilities to deal with fuzzy terms. Its structures and operations are based on fuzzy set theory and its operations. Adaptive fuzzy logic systems are fuzzy logic systems equipped with training algorithms so that its rules can be extracted or modified from available numerical data similar to neural networks. Both fuzzy logic systems and multilayer perceptrons have been proved to be universal function approximators. Since there are approximations in almost every stage, both of these system types are good candidates for classification systems. In classification problems unequal learning of each class is normally encountered. This unequal learning may come from different learning difficulties and/or unequal numbers of training data from each class. The classifier tends to classify better for a well-learned class while doing poorly for other classes. Classification costs that may be different from class to class can be used to train and test a classifier. An error backpropagation algorithm can be modified so that the classification costs along with unequal learning factors can be used to control classifier learning during its training phase. Ph. D.

A digital controller was designed for a semiactive primary suspension for a class 8 highway truck. The controller used a skyhook policy (where the semiactive damper simulates a damper between the sprung mass and an inertial reference) to control magneto-rheological dampers placed on the truck 's primary suspension in response to measurements made by accelerometers placed on the axle and the truck frame. The completed system was then tested for both random noise (on highway driving) and impulse (speed bump) response. The test results showed that for the damping tuning and controller arrangements used in this study, semiactive dampers do not offer any significant benefits in reducing overall vibration levels at the truck frame or axles. The semiactive dampers, however, provided better control of the dynamic transients, such as roll and pitch induced by hitting speed bumps, as compared to passive dampers. Further assessment of the magneto-rheological damper's tuning and the skyhook control policy is needed to establish any definitive conclusions on the potential benefits of semiactive magneto-rheological suspensions for heavy trucks. Master of Science

A field-oriented control design and detailed analysis of a VSI-fed PMSM drive system with active load is done through simulations of the system model, using modern simulation software packages. A new control method for the speed tracking control based on the estimation of the load torque profile is proposed. A new, multilevel modeling approach for creating simulation models of power electronic circuits is developed for easier analysis and faster simulations. It is based on a modular approach wherein each module can be modeled at any level of complexity, while maintaining full compatibility of the modules. The new approach is applied to modeling of the VSI-fed PMSM drive system. The three-phase VSI-fed PMSM drive system model that is developed and experimentally verified is analyzed in the application of a starter/generator, where the load changes dynamically with motor speed. As a result, a detailed analysis of the field-oriented control design of a two stage cascade digital controller is presented, with an emphasis on the new method for the speed control, large-signal and small-signal analyses of several most popular flux-weakening strategies, and sampling delay effects on the system stability. Master of Science

The Interactive Video Data System (IVDS) project began with an initial abstract concept of achieving interactive television by transmitting hidden digital information in the audio of commercials. Over the course of three years such a communication method was successfully developed, the hardware systems to realize the application were designed and built, and several full-scale field tests were conducted. The novel coding scheme satisfies all of the design constraints imposed by the project sponsors. By taking advantage of psychoacoustic properties, the hidden digital signature is inaudible to most human observers yet is detectable by the hardware decoder. The communication method is also robust against most extraneous room noise as well as the wow and flutter of videotape machines. The hardware systems designed for the application have been tested and work as intended. A triple-stage audio amplifier buffers the input signal, eliminates low frequency interference such as human voices, and boosts the filtered result to an appropriate level. A codec samples the filtered and amplified audio, and feeds it into the digital signal processor. The DSP, after applying a pre-emphasis and compensation filter, performs the data extraction by calculating FFTs, compensating for frequency shifts, estimating the digital signature, and verifying the result via a cyclic redundancy check. It then takes action appropriate for the command specified in the digital signature. If necessary it will verbally prompt and provide information to the user, and will decode infrared signals from a remote control. The results of interactions are transmitted by radio frequency spread spectrum to a cell cite, where they are then forwarded to the host computer. Master of Science

This work is a study of a hybrid adaptive controller that blends fixed feedback control and adaptive feedback control techniques. This type of adaptive controller removes the requirement that information about the disturbance is known apriori. Additionally, the control structure is implemented in such a way that as long as the adaptive controller is stable during adaptation, the system consisting of the controller and plant remain stable. The objective is to design and implement an adaptive controller that damps the structural vibrations induced in a multi-modal structure. The adaptive controller utilizes an adaptive infinite impulse response lattice filter for improved damping over the fixed feedback controller alone. An adaptive finite impulse response LMS filter is also implemented for comparison of the ability for both algorithms to reject harmonic, narrow bandwidth and wide bandwidth disturbances. It is demonstrated that the lattice filter algorithm performs slightly better than the LMS filter algorithm in all three disturbance cases. The lattice filter also requires less than half the order of the LMS filter to get the same performance. Master of Science

Flexible manufacturing systems (FMS) are production systems consisting of identical multipurpose numerically controlled machines (workstations), automated material handling system, tools, load and unload stations, inspection stations, storage areas and a hierarchical control system. The latter has the task of coordinating and integrating all the components of the whole system for automatic operations. A particular characteristic of FMSs is their complexity along with the difficulties in building analytical models that capture the system in all its important aspects. Thus optimal control strategies, or at least good ones, are hard to find and the full potential of manufacturing systems is not completely exploited. The complexity of these systems induces a division of the control approaches based on the time frame they are referred to: long, medium and short term. This thesis addresses the short-term control of a FMS. The objective is to define control strategies, based on system state feedback, that fully exploit the flexibility built into those systems. Difficulties arise since the metrics that have to be minimized are often conflicting and some kind of trade-offs must be made using "common sense". The problem constraints are often expressed in a rigid and "crisp" way while their nature is more "fuzzy" and the search for an analytical optimum does not always reflect production needs. Indeed, practical and production oriented approaches are more geared toward a good and robust solution. This thesis addresses the above mentioned problems proposing a fuzzy scheduler and a reinforcement-learning approach to tune its parameters. The learning procedure is based on evolutionary programming techniques and uses a performance index that contains the degree of satisfaction of multiple and possibly conflicting objectives. This approach addresses the design of the controller by means of language directives coming from the management, thus not requiring any particular interface between management and designers. The performances of the fuzzy scheduler are then compared to those of commonly used heuristic rules. The results show some improvement offered by fuzzy techniques in scheduling that, along with ease of design, make their applicability promising. Moreover, fuzzy techniques are effective in reducing system congestion as is also shown by slower performance degradation than heuristics for decreasing inter- arrival time of orders. Finally, the proposed paradigm could be extended for on-line adaptation of the scheduler, thus fully responding to the flexibility needs of FMSs. Master of Science

This dissertation presents new fuzzy logic techniques for designing control systems for a wide class of complex systems. The methods are developed in detail for a crane system which contains one rigid-body and one oscillation mode. The crane problem is to transfer the rigid body a given distance such that the pendulation of the oscillation mode is regulated at the final time using a single control input. The investigations include in-depth studies of the time-optimal crane control problem as an integral part of the work. The main contributions of this study are: (1) Development of rule-based systems (both fuzzy and crisp) for the design of optimal controllers. This development involves control variable parametrization, rule derivation with parameter perturbation methods, and the design of rule based controllers, which can be combined with model-based feedback control methods. (2) A thorough investigation and analysis of the solutions for time-optimal control problems of oscillation mode systems, with particular emphasis on the use of phase-plane interpretation. (3) Development of fuzzy logic control system methodology using expert rules obtained through energy reducing considerations. In addition, dual mode control is a "spin-off" design method which, although no longer time optimal, can be viewed as a near-optimal control method which may be easier to implement. In both types of design optimization of the fuzzy logic controller can be used to improve performance. Ph. D.