Fuzzy neural networks combining advantages of numerical computations of neural networks and symbolic processing originating from artificial intelligence are typical constructs of computational intelligence. The paper analyzes basic processing components (fuzzy neurons), proposes several general architectures, elaborates on learning algorithms, and provides a series of application cases.

The ability to predict failures in machinery before they occur would save time, money, and lives. The Army has several areas that would benefit from this ability. Mechanical components could be replaced before they caused catastrophic damage. Electronic components could be replaced in communication and weapon systems before they endangered a mission or lives. One area that would benefit immediately from this ability is predicting the fatique life of the Army's CH-47 helicopter. The CH-47 is a twin-rotor platform that depends on the reliability of its engine, transmissions, rotors, flight controls, and a myriad of other equipment. Predicting the fatigue life of a CH-47 would save the Army operation and support costs through spares elimination and more timely maintenance cycles. We have developed a methodology for a machine fatigue life predictor that utilizes a combination of parameter estimation, model generation, and condition identification. Using data collected from various fault conditions on the tail rotor assembly of a helicopter, we have simulated fatigue conditions and demonstrated the developed methodology.

Segmentation of images is used for several purposes such as estimation of the boundary of an object, shape analysis, contour detection, texture segmentation, and classification of objects within an image. Despite the existence of several methods and techniques for segmenting images, this task still remains a crucial problem. In our research we have developed a neural network-based fuzzy clustering technique to segment images into regions of specific interest using a quadtree segmentation approach. Since different regions of an image contain varying amount of detail, it is advantageous to segment the regions into blocks of different sizes depending on the content of information present within each block. As the global features of an image are distributed over a wider span of the image and the finer details are concentrated in limited regions, a quadtree segmentation algorithm can efficiently tackle the problem of segmenting images of all kinds. However, block-based techniques tend to introduce blocking artifacts and this problem can be avoided by using a neuro-fuzzy clustering scheme to merge the neighboring blocks of similar regions in a smooth fashion. The proposed algorithm has been applied to images of different kinds and has yielded promising results.

This paper illustrates the neural net approach to constructing a fuzzy logic decision system. This technique employs an artificial neural network (ANN) to recognize the relationships that exist between the various inputs and outputs. An ANN is constructed based on the variable present in the application. The network is trained and tested. After successful testing, the ANN is exposed to new data and the results are grouped into fuzzy membership sets. This data grouping forms the basis of a new ANN. The network is now trained and tested with the fuzzy membership data. New data is presented to the trained network and the results from the fuzzy implications. This approach is used to compute skid resistance values from G-analyst accelerometer readings on open grid bridge decks.

A generalized controller based on fuzzy clustering and fuzzy generalized predictive control has been developed for nonlinear systems. The identification of the system to be controlled is realized by a three-layer feed-forward neural network model. The speed of convergence of the neural network is improved by the introduction of a fuzzy logic controlled backpropagation learning algorithm. The neural network model is then used as the system simulator for developing the predictive fuzzy logic controller. The use of fuzzy clustering facilitates automatic generation of membership relations of the input-output data. Unlike the linguistic fuzzy logic controller which requires approximate knowledge of the shape and the numbers of the membership functions in the unput and output universes of the discourse, this integrated neuro-fuzzy approach allows one to find the fuzzy relations and the membership functions more accurately. Furthermore, there is no need for tuning the controller. A nonlinear heating/cooling system is chosen as an application. The performance of this predictive fuzzy controller is shown to be superior to that of on/off, PI, and linguistic fuzzy logic controllers in terms of both the accuracy and the consumption of energy.

Emerging approaches to modeling fuzzy neurocomputing are solely based on logic operators and/or represented by min and max operators and their extensions--triangular norms. The max-min based models of fuzzy neurons suffer from a kind of insensitivity: only the weakest (strongest) argument(s) affects the result of min (max) operator. While this feature could be taken as an advantage in some cases, in genral it may produce highly undesireable results. Aggregating inputs of a neuron with max-min model results in ignoring most of incoming pieces of information. On the other hand, replacing max-min operations by triangular norms, though removing insensitivity drawback, creates a problem related to a relevant handling of negative nature of information to be processed in neural network. In this paper, new fuzzy-set oriented model of neuron is introduced and analyzed. The architecture of this neuron is based on the selection of positive and negative types of information. The idea of such selection was applied in fuzzy neurons introduced by Hirota and Pedrycz and Pedrycz and Rocha. While the models of those neurons involve max and min operators and triangular norms, the neuron presented here utilizes a certain extension of fuzzy sets. This extension is based on algebraic operators rather than on triangular norms. Subsequently, the formalism is capable of representing negative as well as repetitive information--an evident advantage over conventional fuzzy set connectives. Moreover, processing fuzzy information with aglebraic operators is compatible with the nature of common models of nonfuzzy neurons. Main features of introduced models of neuron are presented and some characteristics of the neuron are discussed. The comparison with other models of fuzzy neurons is also summarized.

In this paper we extend previous work done at the NASA Goddard Space Flight Center on the development of a fuzzy logic controller to generate individual joint rate commands to the robotic arm in response to human input, via remote manipulation of translational and rotational hand controllers. In this phase of the study we have utilized the Goddard Space FLight Center Remote Manipulator hardware/software system in which the fuzzy logic controller has currently been implemented, and have used this system for tuning the controller for near optimal response. Performance results from the system studies are presented.

We have extended the least biased fuzzy clustering algorithm to inhomogeneous data sets. The resolution parameter is generalized from a scalar to a vector with the dimension of the feature space. We fix the orientation of the resolution vector to measure the relative inhomogeneities of each cluster of data points in the different dimensions; and we study the effect of the magnitude of the resolution parameter on the phase transitions yielding the clusters. Based on the detection of the onset of a phase transition, a new technique for truncating the iteration scheme of solution reduces the computational complexity to the order of the number of data points. The actual computational load of the algorithm is discussed and examples are given to illustrate the performance of the algorithm in clustering inhomogeneous data sets.

Two applications of fuzzy integrals to handwritten word recognition are discribed. Fuzzy integrals are used in a dynamic programming based, offline handwritten word recognition algorithm. This algorithm finds optimal matches between word images and strings in lexicons. Fuzzy integrals are used to compute the match scores. Fuzzy integrals are also used in a hidden Markov model based, offline handwritten word recognition algorithm. In this case, fuzzy measures and integrals are used as alternatives to probabilistic measures and ordinary integrals. Results are presented on standard data sets consisting of real images collected form United States Postal Service mail stream.

We will provide a brief description of the field of approximate reasoning systems, with a particular emphasis on the development of fuzzy logic control (FLC). FLC technology has drastically reduced the development time and deployment cost for the synthesis of nonlinear controllers for dynamic systems. As a result we have experienced an increased number of FLC applications. In a recently published paper we have illustrated some of our efforts in FLC technology transfer, covering projects in turboshaft aircraft engine control, stream turbine startup, steam turbine cycling optimization, resonant converter power supply control, and data-induced modeling of the nonlinear relationship between process variable in a rolling mill stand. These applications will be illustrated in the oral presentation. In this paper, we will compare these applications in a cost/complexity framework, and examine the driving factors that led to the use of FLCs in each application. We will emphasize the role of fuzzy logic in developing supervisory controllers and in maintaining explicit the tradeoff criteria used to manage multiple control strategies. Finally, we will describe some of our FLC technology research efforts in automatic rule base tuning and generation, leading to a suite of programs for reinforcement learning, supervised learning, genetic algorithms, steepest descent algorithms, and rule clustering.

We present an extension of fuzzy controllers to include multimedia rules, i.e., rules which do not include verbal or numerical descriptors. We describe the structure and construction of such a multimedia fuzzy controller. In particular, we describe an empirical but unbiased methodology to measure, from human subjects, distances in feature space and hence determine fuzzy memberships. We also propose a practical multimedia fuzzy controller and describe its application examples are given from the law enforcement field where man-machine interactions are important and applications of the methodology described in this paper appear promising.

In this paper, we discuss development of useful intelligent systems, emphasizing an integrated systems perspective. In particular we will highlight two important ideas. The first is associated with the sensor-based control, for realizing effective and robust approaches for processing and analysis of sensory information associated with the state of the environment and the state of the intelligent system for reflexive and deliberative tasks. The second is associated with cooperation, for developing teams consisting of heterogenerous and complementary intelligent modules. The aim is to develop useful and effective intelligent systems involving teams of players which are either all robotic devices or some robotic and some human.

The Earth's surface and subsurface can be mapped with a very high resolution using synthetic aperture radar (SAR). It is an effective tool for remote sensing, since it works nearly independent from daylight and weather conditions. However, the huge amount of signal data difficults the on-board storage and the data downlink to the ground station. This paper presents a raw data compression algorithm using fuzzy logic, the fuzzy block adaptive quantizer (FBAQ), which improves the efficiency of the system. The data rate is reduced by a factor of 3 with a slight deterioration of the image quality. The performance of the FBAQ is compared with the traditional algorithm, a block adaptive quantizer and its hardware design for implementation in an airborne SAR system is shown.

A fuzzy logic preprocessing is used in connection with a back propagation neural network in particle recognition. As application on 4 GeV CERN experimental and Monte Carlo data of e^- and (pi) (superscript -, taken with the prototype of the silicon Tungsten calorimeter of the Wizard collaboration, is shown. This preprocessing consists in giving as input to the net the membership value, for a given discriminating parameter value, to belong to a given particle class. In this way the input layer receives a normalized input. The net can then exploit the correlations between different parameters, resulting in an increased convergence speed and recognition capability of the net. Other advantages of this approach are its noise robustness and the simple generalization of other particle classes or energies.

Learning vector quantization (LVQ) often requires extensive experimentation with the learning rate distribution and update neighborhood used during iteration towards good prototypes. A single winner prototype controls the updates. This paper discusses two soft relatives of LVQ: the soft competition scheme (SCS) of Yair et al. and fuzzy LVQ equals FLVQ. These algorithms both extend the update rates that are partially based on posterior probabilities. FLVQ is a batch algorithm whose learning rates are derived from fuzzy memberships. We show several relationships between SCS and FLVQ; and we show that SCS learning rates can be interpreted in terms of statistical decision theory. Finally, we show the relationship between FLVQ, fuzzy c-means, hard c-means, a batch version of LVQ, and SCS.

The automation of chromosome identification and visualization for a complete cell (karyotyping) has been the subject of considerable research. While rather high classification rates are possible on individual chromosomes, the cell level classification rates are still quite low. We describe a system which uses partial confidence values generated by neural and fuzzy classifiers with optimization to increase the cell level recognition rates. This is consistent with Marr's Principle of Least Commitment for the design of intelligent computer vision algorithms.

In many applications, the use of multiple robots of different capabilites will be desirable. In a teaming approach, multiple robots will cooperate to accomplish a task which a single robot alone is incapable, inefficient, or impractical at performing. When multiple robots are deployed, the motion of each robot may have to be coordinated with the rest of the robots. This paper deals with development of a fuzzy logic based motion controller for convoying the coordinated motion between a leader and a follower. The fuzzy logic based real-time motion controller uses relative distance and speed information to adust the speed of the trailing robot to smoothly follow the leader without either collisions or loosing sight of it. The performance and robustness of the system has been verified in a series of extensive laboratory trials.

Object recognition is usually obtained by using some kind of different approach based on profile reconstruction, whose results are not always reliable or sufficient for a right object identification. The present paper describes an alternative method to profile calculation, based on color and surface information associated to the images. The method is also used to define the system typology on the basis of a detailed analysis of human instinctive behaviors associated to the task to be performed by the system.

This paper presents the development of generalized fuzzy k-means algorithms and their application in image compression based on vector quantization. The development of generalized fuzzy k-means algorithms is based on the search for partitions of the feature vector space other than those generated by existing fuzzy k-means algorithms. These alternative partitions can be obtained by relaxing one of the conditions imposed on the membership functions. The clustering problem is formulated as a constrained minimization problem, whose solution depends on the selection of a constrain function that satisfies certain conditions. The solution of this minimization problem results in a broad family of generalized fuzzy k-means algorithms, which include the existing fuzzy k-means algorithms as a special case. Moreover, the proposed formulation results in the minimum fuzzy k-means algorithms, which are computationally less demanding than the existing fuzzy k-means algorithms. A broad family of admissible constrain functions result in an extended family of fuzzy k-means algorithms, which at the limit provide the fuzzy k-means and minimum fuzzy k-means algorithms. The resulting algorithms are used in image compression based on vector quantization.

Object recognition is often performed in an environment full of uncertainties. Typical factors are the imprecision introduced by the limitations of image processing algorithms, the misinterpretation of the feature vector due to noise or occlusion, and the infinite variability of the object features due to continuous environment change as well as countermeasures. An integrated approach (statistical methods, multisensor fusion, and fuzzy logic) for automatic object recognition if presented in this paper. A fuzzy scene representation is proposed to cope with uncertainties. The features of the object and the background are obtained from both a priori knowledge and the data collected by a multisensor suite and then reconstructed for object recognition. A recognition scheme, based on fuzzy logic, has been developed to merge the information from multiple sources of differing resolution and confidence into a combined assessment of the object identity. It has been found that the fuzzy logic based information fusion architecture provides a platform to accommodate the output of different types of image/signal processing algorithms. In addition, it allows the input of the temporal scene development through an expanded feature vector. Details of the fuzzy scene representation and the recognition process are discussed. Experimental results are presented to show the potential of the approach.

An approach to uncertainty and decision making problems alternative to probabilistic one is studied. It is based on fuzzy sets technique and has deep relations with the underlying first order fuzzy logic. In this logic not only logical connectives but also quantifiers have fuzzy interpretation. It is shown that all fundamental concepts of probability and statistics such as joint distribution, conditional distribution, etc., have meaningful analogs in new context. Thus the classical concepts obtain fuzzy-logical semantics. As a result one may treat them as formulas of first order logic. It is shown, that this appraoch makes it possible to utilize rich conceptual experience of statistics. In particular it leads to fuzzy Bayesian approach in decision making, which plays the same role in fuzzy problems of optimal decision making as its probabilistic prototype in the theory of statistical games, and provides methods for construction of optimal strategies. Connection with underlying fuzzy logic provides the logical semantics for fuzzy decision making. In this approach a priori information is represented by a fuzzy predicate and an experiment--by fuzzy universal quantifier, etc. As a result the notion 'good decision strategy' is expressed by a first order formula in this logic.

The membership functions are the critical factors of the system performance for fuzzy logic control. The degree of membership is determined by a membership function, which is defined based on experience or intuition of engineers. In general, it is accepted that membership functions changes several times as the system is tuned to achieve desired responses to given inputs. Once the system is in operation, the membership functions do not change. So the membership functions must be tuned properly to obtain the desired system performance. In this paper, we propose the technique to overcome the tuning difficulties of membership functions by trial and error method. This technique has been applied to the system with parameter uncertainties to investigate the performance.

In this paper, a TSK type fuzzy logic controller (FLC) with sliding mode controllers (SMC) as its rule consequents is developed, namely a fuzzy sliding controller (FSC). As FSC is a nonlinear controller and its parameters can be easily decided through an analysis of the desired slope and the boundary layer's thickness of the SMC in different regions in the state space. Moreover, the FSC can provide a nonlinear static mapping with fewer rules, and its stability can be analyzed through the sliding condition of the SMC. The derivation of gradient descent based update equations for learning the parameters in each rule of the FSC is included in this paper. Simulation results with an inverted pendulum model are presented. Comparisons between the FSC and a TSK type FLC with linear output functions are also included in this paper.

A ball balancing beam is a nonlinear dynamic system which is quite difficult to control using convention methods since some special mathematical techniques and control theory knowledge are required to derive the equations. There are difficult issues in this system: it has delayed feedback associated with control actions; and the 'jumping ball' phenomenon brings sensor uncertainty. Balancing is essential in carrying out robotic tasks such as transporting dynamic system. Relative to conventional controllers, a fuzzy logic controller requires less mathematical derivation in design; it has high noise tolerance, probably gravity independence, and it is portable to other scales of same system setup such as with a shorter beam or a heavier ball. In this paper, membership function construction, and fuzzy rule generation are discussed in detail. The hardware setup is shown. In addition, ways to overcome feedback delays and noisy signals are presented. Finally, the performance of the fuzzy controller will be evaluated and compared to conventional and neural network controllers.

A systematic approach is presented for modeling qualitative properties in semiconductor manufacturing processes. This approach is based on the fuzzy logic theory, and on the statistical analysis of categorical data. A fuzzy inference system can be designed and created by training data obtained either from human expert knowledge, or automatically extracted from statistically designed experiments. Before being used to design the fuzzy system, the data extracted from the designed experiments can be processed and filtered with the help of linear and logistic regression analysis. After the establishment of the initial inference system, the fuzzy membership functions can be tuned adaptively to accommodate process changes.

This paper discusses the potential of fuzzy logic for manufacturing science. After a general introduction of manufacturing science as a complex systems to which Zadeh's Incompatibitlity Principle applies, the paper presents a bird's eye view of fuzzy logic applications to operational and engineering aspects of manufacturing science, and, in particular, to reliability studies.

In this paper, an intelligent control architecture has been proposed to model, monitor, and control the product quality in a CNC milling machining center. The adopted architecture uses a fuzzy timed Petri net (fuzzy logic with timed Petri nets) and synergistic neural networks. The paper demonstrates how fuzzy input variables, fuzzy marking, fuzzy time-firing sequences, fuzzy outputs reasoning, and a global output variable should be defined for use with fuzzy timed Petri nets. The technique employed with the fuzzy timed Petri net uses two fuzzy input variables, spindle speed, and feed rate, through the milling process in order to maintain product quality characteristics such as surface finish. It also illustrates how the technique reacts when the product quality is high, medium, or low.

Within the context of our sensor fusion systems, we define an entity's vulnerability as the certainty with which other entities have the capability to detect and/or strike the entity; vulnerability assessment (VA) is the inference of vulnerability certainties. This investigation considers two issues: the feasibility of a fuzzy VA algorithm and the interface of a fuzzy VA algorithm into an existing sensor fusion system, including human-machine interface aspects. Relative kinematics, sensor/weapon technical capabilities, sensor/weapon system state, contextual electronic signatures, physics, terrain, atmospherics, and doctrinal bias are certainly all viable inputs to a VA algorithm. These data are traditionally characterized by a mix of continuous, discrete, and/or symbolic values with associated error bounds in various mathematical forms. Hence, the algorithmic infusion of a fuzzy VA into this systemic environment implies resolving the uncertainty information content of these representations and integrating them into a coherent fuzzy reasoning context. The information overload facing the tactical operator has necessitated the reduction of many data to prioritized simple alerts. While there is a reasonable understanding of the visual representations and implications of thresholding probabilistic data, the presentation and thresholding of fuzzy data is not well understood; some of the more critical implications on the human-machine interface are presented herein.

An interactive query specification method in fuzzy document retrieval using a genetic algorithm is proposed based on a possible-worlds perspective where every world is represented by a bit code of 1 and 0 which corresponds to truth and false, respectively. Given a query, several seed worlds are generated in a set of worlds and seed worlds. Schemata generated through selection, crossover, and mutation can be regarded as candidates for possible needs implicit in the user, and then he can select some of these schemata to specify his need and revise his query interactively.

The paper presents the inverse problem concerned with fuzzy relations. A method for the inverse solution of a max-min compositional relation equation, based on the intermediate value theorem, is called the bisection algorithm, or binary-search method. The benefits of the inverse problem is used not only for its theoretical implications but also for its applicability on real problems of optimization and control of fuzzy logic systems.

This paper describes a stand alone Mountain Method (MM) to determine clusters' fine center position, extents, and membership functions. The motivation for this fuzzy clustering algorithm is to improve the approximate cluster centers resulting from the MM without experiencing the problem of local minima, and provide the additional cluster information necessary to fuzzify a system. After briefly reviewing the MM, this paper discusses the methodological considerations for effective cluster determination, modifications to the MM, and classical data set results using the MM, fuzzy c-mean (FCM) and the modified Mountain Method (MMM). The MMM compared to FCM algorithm achieved cluster centers resulting in a standard error of 0.166 and 0.052 with respect to Anderson's Iris, and Yager and Filev's data sets. The methodology discussed herein is the subject of a US patent, entitled 'Multiple Target Discrimination Algorithm', serial number 5,396,252, assigned to United Technologies Corporation. The patent addresses discriminating multiple targets or clusters of digitized sensor data for determining fine position and extents thereof.

A new fuzzy learning and classification scheme based on developing a fuzzy understanding of neighborhoods with nearest unlike neighbor sets (NUNS) is reported in this study. NUNS, by definition, the set of samples identified as the nearest from the other class(es) for each given sample, represent in essence the boundaries between pattern classes known in the problem environment. Accordingly, samples close to the NUNS are likely to have more ambiguity or uncertainty in their labels than those farther away from these NUNS. This information about the uncertainty or imprecision in the labels of the given training set can be extracted and represented in terms of fuzzy memberships. These fuzzy membership values, which may be determined in the learning phase using appropriate fuzzy membership models, can then be utilized in the classification phase to derive the identity of an unknown sample. This classification can be accomplished using any one of the established fuzzy classification techniques.

A fuzzy logic (FL) anti-skid brake controller (ABS) is proposed as the next generation ABS replacing current generation finite state (FS) control. The FL controller is part of a commercial truck braking system, encompassing reverse front-back braking proportions on an articulated vehicle as compared to that found on fixed, passenger car systems. In this early research, the FL controller must satisfy three goals. The first goal is to produce superior braking distances over that of the finite state controller, specifically under low (mu) conditions. The second goal is to provide superior braking under varying system conditions (road surface conditions, physical brake parameters, wheel velocity sensor parameters). The third goal is to provide a convenient, flexible, and tractable ABS solution which is amenable to redevelopemnt to different vehicular platforms. Monte Carlo simulation results illustrate stopping distance improvements of 5 to 10 % averaged over all (mu) surfaces for varying wheel loads. On low (mu) surfaces, the improvement increases to 15% (up to a full tractor-trailer length). These results are obtained while varying other system parameters demonstrating robustness. Finally, the fuzzy logic rule sets and the overall configuration illustrate a straight-forward design and maturation process for the rule sets.

In this work the principles of making a method for creating an object with required attributes are discussed. The method was illustrated by the example of the creation of analytical function, which has a geometrical image closely spaced to a required one. The computer program was made to implement the discussed method. The convergence of the algorithm was analyzed.

Human perception of pictorial visual information is investigated from iconical sign view-point and appropriate semiotical model is discussed. Image construction (syntactics) is analyzed as a complex hierarchical system and various types of pictorial objects, their relations, regular configurations are represented, studied, and modeled. Relations between image syntactics, its semantics, and pragmatics is investigated. Research results application to the problems of thematic interpretation of Earth surface remote imgages is illustrated.

Referring to the different feedback information, the structure--changeable fuzzy controller presented in this paper select different control modes, such as PID, FLC, separately. It is known by studying the simulation results that the structure--changeable fuzzy controller, can eliminate the steady state error and result better robustness than classic PID controller.