To assist the battlefield commander, the U.S. Army Atmospheric Sciences Laboratory has developed a rule-based expert system that can advise weather forecasters on the possibility of thunderstorm activity and other severe weather phenomena. This expert system, the Thunderstorm Intelligence Prediction System (TIPS), provides expedient conclusions about the weather that are derived from a large sum of data. During the spring and summer thunderstorm season of 1990 TIPS was tested at 14 different sites across the United States

Today, over hundred tools (or shells) for building expert systems are commercially available. However, most of them are not generally appropriate for building active expert systems. This paper presents a tool called BUOSHI for developing such expert systems. The flexibility for handling guaranteed response time is the key capacity of the tool.

This paper describes the functions, the advantages, and the performing steps of preventive control in power plants, and presents the methods to construct this prototype expert system. The differences between preventive control and diagnosis are also described. Preventive control consists of two major areas: (1) Correction of process control of generator system, and (2) prevention of equipment failure. This system uses frames along with rules to construct a large knowledge base. there are three main methods for developing the knowledge base; (1) Constructing root frame and subframes according to the properties of operational process and equipment; (2) determining all parameters used in the knowledge base; and (3) turning the expertise into rules by means of decision trees. Examples are presented for describing this prototype expert system.

Scientists at the U.S. Bureau of Mines are currently investigating ways to combine the control capabilities of fuzzy logic with the learning capabilities of genetic algorithms. Fuzzy logic affords a mechanism for incorporating the uncertainty inherent in most control problems into conventional expert systems. Although fuzzy logic-based expert systems have been used successfully for controlling a number of physical systems, the selection of acceptable fuzzy membership functions has generally been a subjective and time consuming decision. In this paper, high-performance fuzzy membership functions for a fuzzy logic controller that manipulates a mathematical model simulating a cart-pole balancing system are selected using a genetic algorithm, a search technique based on the mechanics of natural genetics. The membership functions chosen by the genetic algorithm provide for a more efficient fuzzy logic controller than membership functions selected by the author for the cart-pole balancing problem. Thus, genetic algorithms represent a potentially effective and structured approach for designing fuzzy logic controllers.

In the so called 'Group Problem' [4] (which we term as 'Group- Consulting Type Cooperation'), ESs of the same cooperating group may entertain different views towards the same problem, and the differences may sometimes be very great. For the circumstance of these ESs entertaining relatively closer views, we present three methods for consistency treatment, and a method for calculating the maximal deviation of the confidence value about the cooperating group's common view based on the consistent degree of those different views. For the case of those views being not very close to and also not very different from each other, we designed two synthesis functions for synthesizing multiple views. (These circumstances are formally specified). These two functions are also compared with the combining function of [4].

Object detection approaches need to perform accurately and robustly over a wide range of scenes. It would be quite valuable if one can devise a performance index for an object detection approach as a function of the nature of a particular scene. Basically this requires an ability to derive a quantitative measure for the 'clutter' observed in an image. Most images of interest are texture-rich i.e. the important perceptual properties are based upon the spatial arrangements of simple patterns which might be regular in nature. As a result, it is natural to utilize texture analysis based operators to define the measure of image quality of 'clutter' that is being sought. It has been proven that the gray level cooccurence (GLC) matrices of an image embody important texture information, and the image can indeed be reconstructed from these matrices. Hence it is proposed that GLC-based measures be derived and used to quantify image quality. Current approaches are based on only one of several important perceptually meaningful measures which can be computed from GLC matrices. Prior work done in this area is assessed in this paper. The derivation of the image quality measures from GLC matrices is currently being researched. This paper presents a discussion of these issues along with the objectives and results of an ongoing study involving object detection in high resolution aerial images.

The foremost difficulty in the Multiple Target Tracking (MTT) field involves the problem of associating the measurements with the tracks when there are missing reports and the proliferation of false reports generated by clutter. In this paper, in order to solve this return-to- track association as-well-as the estimation problem, we are proposing to associate Artificial Intelligence techniques and Signal Processing method. The Signal Processing method is the Multiple Hypothesis Filter (M.H.F.).

A typical hierarchy for a general object recognition problem consists of object detection, classification and identification. Detection pinpoints the presence of an object or objects, classification categorizes the object(s), and identification distinguishes the object(s). This paper establishes necessary building blocks required for high-speed object recognition applications. An architecture that combines digital and optical processing, exploiting current image processing techniques for detection and classification, and optical processing hardware is described. An optical processing scheme is suggested for the identification aspect. Pre-processings that suppress background noise, minimize the number of matching filters and optimize post-processings of correlation outputs are performed by initially detecting objects in a background suppressed 2D scene via texture analysis and blob representation (detection), then scale/rotation estimation and shape recognition techniques (classification) are used as a precursor to optical processing. Post processing techniques which analyze and detect correlation peak(s) are also discussed. In addition, numerical results of each proposed concept are presented.

A typical object recognition problem deals mainly with the characterization of objects and treats background as noise. However, in many applications, we can assume the reverse. Intuitively, the background information is easier to describe in terms of intensity distributions (textures), while the unknown nature of objects' types and positions poses a more difficult formulation for general pattern recognition problems. In this paper, we propose the use of a second order moment function to describe the scene. Various types of backgrounds and objects are evaluated, and results from this experiment are presented.

When a human being doesn't know how to solve a problem, one of the methods at his disposal consists in using problem-solving knowledge on easier problems. This knowledge can then serve as heuristics to guide the search for a solution to the original problem. Many AI researchers have already profited from this idea and in our turn we have adopted it to seek solutions to workshop scheduling problems. Setting up scheduling to satisfy, as well as is possible, the preferences that have been expressed is very complicated when the set of constraints is too large and conflictual to be solved directly. The strategy we propose consists in evaluating the interaction between constraints and grouping those that lead to the same schedule into simplified subproblems. A solution to these subproblems takes the form of a series of tasks satisfying the constraints. In combination with others, this solution can guide the scheduling system in its choices for constraint satisfaction. Establishing such a strategy requires close cooperation between constraint analysis decisions and scheduling decisions. But maximum effectiveness precludes any rigidly predefined way of organizing this cooperation. Thus the system must be able to adapt its problem-solving strategy in terms of evolution in the solution. The cooperative and opportunistic nature of the system has led us to choose a 'blackboard' based architecture.

The objective of this project was to use expert system technology to aid in the scheduling activities performed at the White Sands Missile Range (WSMR). The WSMR range scheduling problem presents a complex interactive environment. A human factors approach was undertaken, in that, the goal was to implement a system which mimics current WSMR scheduling procedures. The results of this project have produced a prototypic scheduling tool, called Scheduler's Assistant (SA), to aid WSMR range schedulers to generate a daily schedule. The system provides resource conflict detection and resolution advice through a series of cooperating expert systems. Immediate advantages of the system are increased safety, insurance of proper schedule execution and improved speed for turnaround time of sudden schedule changes. Additional benefits of SA include: expandability as future operations grow, allows for rapid redeployment for changing resources, promotes efficient management of WSMR resources, provides a formal representation of knowledge such that years of range personnel experience is preserved and enables the flexibility of a scheduling aid as opposed to a rigid methodology. Prior development efforts by Perceptics have produced a sophisticated expert system development tool, called Knowledge Shaper, which was used to implement all of the expert systems. The development of SA included a library of routines (the SA toolbox) to permit the manipulation of internal data tables and define a data transfer protocol to and from the SA environment. The combination of Knowledge Shaper and the SA toolbox provide a powerful set of design tools for the development of future scheduling applications.

A practical method of reasoning about intervals in a branching-time model which is dense, unbounded, future-branching, without rejoining branches is presented. The discussion is based on heuristic constraint- propagation techniques using the relation algebra of binary temporal relations among the intervals over the branching-time model. This technique has been applied with success to models of intervals over linear time by Allen and others, and is of cubic-time complexity. To extend it to branding-time models, it is necessary to calculate compositions of the relations; thus, the table of compositions for the 'atomic' relations is computed, enabling the rapid determination of the composition of arbitrary relations, expressed as disjunctions or unions of the atomic relations.

In this paper, a general method of active pattern recognition based on attributes and reasoning (AAR) is proposed. The significance of this scheme is to treat active pattern recognition as a planning process of feature based reasoning and to represent the patterns as well as the sources of the patterns accordingly. The planning process of feature based reasoning in active pattern recognition is to examine the potentially 'important' features and to find the optimal feature or a set of features to recognize the source of patterns. This includes the ability to 'look for' the discriminating set of features when there exist some ambiguities among potential sources. This set of features which is being described as 'critical features' allows to collectively reason with symbolic attributes and to avoid unnecessary comparisons of patterns between the unknown source and a series of the model sources from a database.

In many applications, such as remote sensing or target detection, the target objects are small, compact blobs. In the images discussed in this paper these objects are only 6 or fewer pixels across, and the images contain noise and clutter which is similar in appearance to the targets. Since so few pixels comprise an object, the object shape is uncertain, so common shape features are unreliable. To distinguish targets from clutter, features which make use of scale-space have proven useful. The scale-space of an image is a sequence of Gauss-filtered versions of the image, using increasing scales from one image to the next. Experiments show that object features calculated at a single scale. Various moments of the value of the Laplacian at the centroid of a blob were particularly effective for some targets.

This paper describes a set of edge detectors, each of which has selective responses to each edge pattern contained in an input signal. Quantum-Mechanical Harmonic Oscillator Wave Functions are applied to extracting a variety of edge information indispensable to identifying each edge pattern. Information, obtained over the scale space, is also integrated so that edges can be detected even if an input signal is distorted with noise.

Optical flow fields (which describe image domain motion) extracted from sequences of images acquired with moving sensors have many practical applications including motion analysis, moving target cueing, surface reconstruction and hazard avoidance. It is well known that, for the general optical flow problem, a constraint exists that relates the local change in intensity to image gradient and describes the optical flow component parallel to the gradient. However, the component perpendicular to the gradient is unconstrained, and therefore, the optical flow field can not be determined directly-usually it is necessary to appeal to some other constraint, such as smoothness, which generally requires costly iterative or relaxation techniques. For the special, but important, case of optical flow induced by sensor motion, a model for the motion provides additional constraints. If, in addition, the scene is assumed to be roughly planar, the optical flow can be characterized by six parameters and found directly without iteration or determining correspondence. These parameters can be readily and robustly calculated by performing a least squares fit to data uniformly sampled from the image sequence. We develop the necessary equations and relations, show results of this approach using real imagery and demonstrate that it is applicable to important real problems.

One of the fundamental problems in vision is the problem of finding the correspondence between two images. This paper addressed this problem in terms of a computational model consistent with existing models of human vision and human visual psychophysics. At the heart of the model is a multi-dimensional internal representation of the input image. We propose that correspondence may be computed by minimizing the total, non- euclidean distance between corresponding elements. A backpropagation programmed neural network is used to investigate the nature of the internal representation.

An autonomous intelligent agent working in a real, unstructured, dynamic environment must have very close interactions among its perceptual, cognitive, and motor components. We believe that by placing the environment model at the heart of these systems this interaction can be facilitated significantly. In our approach, the environment model is responsible for interaction among different components, providing temporal coherence, combining information from multiple sensors, and providing the purposive behavior to the system. The information about the environment is acquired by using multiple disparate sensors, from multiple viewpoints, and at multiple time instants. We believe that the combination of information from disparate sensors should be viewed as a problem of information assimilation, rather than sensor integration. The focus in information assimilation is on the physical world being modeled, sensory information is just a means to the end. Sensor integration treats the goal implicitly, misplacing the focus on the processing of sensed information. Existing approaches towards autonomous systems tend to follow exclusively reactive or exclusively deliberated operations. We present an approach that provides a balance between reaction and deliberation.

This paper develops a qualitative reasoning methodology for problem solving at multiple levels of abstraction. The goal is to address two important control issues in the behavior generation process: (i) the selection problem, which deals with the right level of detail to solve a problem, and (ii) the efficiency problem, which deals with information transfer from higher levels of abstraction to focus problem solving at more detailed levels. The CMOS digital circuit domain is used as a test bed to illustrate the methodologies developed.

Case Based Reasoning (CBR) is one of the recently emerging paradigms for designing intelligent systems. The preliminary studies indicate that the area is ripe for theoretical advances and innovative applications. Heuristic search is one of the most widely used techniques to solve many real world problems for obtaining optimal solutions. In this paper we identify some necessary properties of the heuristic functions to be solved in the CBR paradigm. We designed a CBR system based on these observations and performed several experiments for a heuristic search problem. We also provide an analysis to compare the performance of the CBR system with the A* search algorithm.

COMIX (Configuring Mixing-Machines) is a knowledge-based system, which was designed to assist members of the distribution department in the configuration of mixing-machines. It's possible to configure a mixing- machine with COMIX with regard to aspects of process-engineering and mechanics or in respect of mechanics only. In the latter case, the user needs to know, what kind of mixing-machine he wants. The knowledge about physical objects, like the components of the mixing-machine, is represented in hierarchies of objects. On the one hand the mixing- machine with it's components takes place in a part-of hierarchy, on the other hand the machine and it's components are represented in different is-a-hierarchies. The knowledge about the process of configuration contains laws of mechanics and process-engineering, as well as industry- standards and heuristics. This knowledge is associated with the frames and instances by functions, relations or rule-sets, which are activated by methods of the objects. The process of configuration is a stepwise refinement of the structural model, until all values of parameters are defined. The process is guided by knowledge about causal dependencies. Alike constraint-propagation, the knowledge about causal dependencies is used, to find the next possible step in the configuration-process, i.e. which parameters can be determined. If there is a re-determination of a parameter with a different value, all values of parameters, which depend on the old parameter value, are retracted, alike a truth maintenance system. The first phase of developing the system is finished and the system will go under a (beta) -test at the company soon.

This paper describes an automation of AEW crew decision processed implemented in an intelligent information system for an advanced AEW aircraft platform. The system utilizes the existing AEW aircraft database and knowledge base such that the database can provide sufficient data to solve the sizable AEW problems. A database management system is recommended for managing the large amount of data. In order to expand a conventional expert system so that is has the capacity to solve the sizable problems, a cooperative model is required to coordinate with five expert systems in the cooperative decision process. The proposed model partitions the traditional knowledge base into a set of disjoint portions which cover the needs of and are shared by the expert systems. Internal communications take place on common shared portions. A cooperative algorithm is required for updating synchronization and concurrent control. The purpose of this paper is to present a cooperative model for enhancing standard rule-based expert systems to make cooperative decision and to superimpose the global knowledge base and database in a more natural fashion. The tools being used for developing the prototype are the ADA programming language and the ORACLE relational database management system.

YAK (Yet Another Krapfen) is a hybrid knowledge representation environment following the tradition of KL-ONE and KRYPTON. In its terminological box (TBOX) concepts and roles are described by means of a language called KFL that, even if inspired to FL-, captures a different set of descriptions, especially allowing the formulation of structured roles, that are aimed to be more adequate for the representational goal. KFL results to have a valid and complete calculus for the notion of subsumption, and a tractable algorithm that realize it is available. In the present paper it is shown how the semantics of a sufficiently significative subset of KFL can be described in terms of standard first order logic semantics. It is so possible to develop a notion of 'relation' between KFL and a full first order logic language formulated ad hoc and usable as formalization of an assertional box (ABOX). We then use this notion to justify the use of the terminological classification algorithm of the TBOX of YAK as part of the deduction machinery of the ABOX. In this way, the whole hybrid environment can take real and consistent advantages both from the TBOX and its tractable classification algorithm, and from the ABOX, and its wider expressive capability.

Successful implementation of sensor based robots in dynamic environments will depend largely upon the immunity of the system to incomplete and erroneous sensory information. This paper introduces a six module, 3D model based robot vision system, which utilizes 3D geometric models of the objects expected to appear in a scene and can tolerate incomplete and noisy image features. Object identification is independent of the particular robot pose and object pose, as long as the object is within view of the camera. The system effectively utilizes topology during the object identification phase to reduce the number of mappings between the domain of image features to that of the object features (object models). Geometric information is then employed by the Pose Determination Module to decipher the identified objects unconstrained position and orientation. Continuing experimentation is giving valuable insight into the characteristics of our strategy and has verified the system performance when using incomplete image feature sets.

A plywood panel is composed of several layers of wood bonded together by glue. The adhesive integrity of the glue formulation employed must surpass the structural integrity of wood used to make the panel. The American Plywood Association (APA) regularly tests plywood from manufacturing plants to ensure that this performance requirement is met. One of the procedures used is to saw a panel into a number of 1 X 3- 1/4 inch blocks called samples. These samples are then subjected to a number of treatments to simulate natural aging. The treated samples are then sheared into two halves. A 1 X 1 inch area on each of the two halves is then visually inspected to determine the percent wood failure that occurred during the shear. Roughly speaking a region of solid wood or a region of wood fibers embedded in glue is considered to be a region of wood failure while a region of glue is considered to be a region of glue failure. If the percent wood failure of sample from a significant number or panels from a plant is too low, the right to use the APA trademarks is withdrawn. The number of samples inspected annually by the APA is in the millions. The human inspectors are well trained, typically having years of experience, and are regularly tested. As in any human endeavor, the inspectors are subject to fatigue, boredom, etc. For these and other reasons an automatic inspection system could aid the APA in better performing its regulatory role.

This paper presents a general vision-based coordinate measurement system, its planned use within a sheet metal inspection application, the expected benefits accrued from the use of an object-oriented design strategy, and the utilization of the methodology as it applies to the redesign of the vision system. Adherence to an object-oriented design strategy will obviate a number of shortcomings present in the original system implementation. System improvements include: a more lucid and intuitive user interface; greater accuracy and reliability through the use of decentralized modules; simplification of the description of the interrelationship between the perceived and actual dimensions of parts being inspected; and, easier porting of the general vision system to specific applications.

This paper presents a new Al approach to general robot vision problems. Specifically, a structure is defined for the inference engine that includes clue-derivation and clue-selection components. The development addressed the role that these components play in the recognition process and the natural learning that occurs in the framework of the clue- selection process.

Nursing diagnosis is an integral part of the nursing process and determines the interventions leading to outcomes for which the nurse is accountable. Diagnoses under the time constraints of modern nursing can benefit from a computer assist. A knowledge-based engineering approach was developed to address these problems. A number of problems were addressed during system design to make the system practical extended beyond capture of knowledge. The issues involved in implementing a professional knowledge base in a clinical setting are discussed. System functions, structure, interfaces, health care environment, and terminology and taxonomy are discussed. An integrated system concept from assessment through intervention and evaluation is outlined.

This paper presents a new design approach to knowledge-based decision support systems for fault diagnosis and control for quality assurance and productivity improvement in automated manufacturing environments. Based on the observed manifestations, the knowledge-based diagnostic system hypothesizes a set of the most plausible disorders by mimicking the reasoning process of a human diagnostician. The data integration technique is designed to generate error-free hierarchical category files. A novel approach to diagnostic problem solving has been proposed by integrating the PADIKS (Pattern-Directed Knowledge-Based System) concept and the symbolic model of diagnostic reasoning based on the categorical causal model. The combination of symbolic causal reasoning and pattern-directed reasoning produces a highly efficient diagnostic procedure and generates a more realistic expert behavior. In addition, three distinctive constraints are designed to further reduce the computational complexity and to eliminate non-plausible hypotheses involved in the multiple disorders problem. The proposed diagnostic mechanism, which consists of three different levels of reasoning operations, significantly reduces the computational complexity in the diagnostic problem with uncertainty by systematically shrinking the hypotheses space. This approach is applied to the test and inspection data collected from a PCB manufacturing operation.

The belief network is a well-known graphical structure for representing independencies in a joint probability distribution. The methods, which perform probabilistic inference in belief networks, often treat the conditional probabilities which are stored in the network as certain values. However, if one takes either a subjectivistic or a limiting frequency approach to probability, one can never be certain of probability values. An algorithm should not only be capable of reporting the probabilities of the outcomes of remaining nodes when other nodes are instantiated; it should also be capable of reporting the uncertainty in these probabilities relative to the uncertainty in the probabilities which are stored in the network. In this paper a method for determining the variances in inferred probabilities is obtained under the assumption that a posterior distribution ont eh uncertainty variables can be approximated by the prior distribution. It is shown that this assumption is plausible if their is a reasonable amount of confidence in the probabilities which are stored in the network.

The paper present the assumptions which were taken into consideration when building an Expert System for Microlithography (ExSyM), and describes this Expert System which has 'learning', 'teaching' and 'answering' functions.

In this paper, a new design of an environment for simulation and animation of sensor-based robots is presented. As sensor technology advances, increasing numbers of robots are equipped with various types of sophisticated sensors. The main goal of creating the visualization environment is to aid the off-line programming capability of sensor- based robots. The software system will help the users visualize the motion and reaction of the sensor-based robot under their control program, thereby increasing the efficiency of program development and reliability of the control program, ensuring the safety of robot operation, and reducing the cost of new software development. Most of the conventional computer-graphics-based robot simulation and animation software packages lack robot sensing simulation capabilities. Our system is designed to overcome this deficiency.

This paper describes the research accomplished by the Industrial Automation Department of CRIF/WTCM in the area of assembly system design. The goal of this project is the integration of the assembly process since the design of the product until the programmation of the assembly cell. The paper presents the structure of the resulting off- line programming system and details two of the most important automatic processes: (1) the assembly planner, and (2) the line design or station allocation. The specific knowledge and the common sense expertise are specified for each module. Several results are presented and discussed on the basis of a concrete case study: a subassembly of a working machine.

In this paper we provide a mathematical support for the nature of the shape representation methods useful for the computation of possible functions of an objects as derived from its shape structure. We discuss the concepts of parts and subparts of objects in the framework of axis based shape representation methods and boundary based methods. We propose a new method for obtaining descriptions of parts which based on a theorem from differential geometry (Pogorelov 1974) that any regular surface can be approximated in a finite environment with a given accuracy by a parabolloid of one of the following types - elliptic, hyperbolic and a parabolloid degenerating into a plane or a parabolic cylinder. Based on these considerations we suggest a heuristic for the approximation of convex object parts by a polyhedra, cylinder, ellipsoid or generalized cone with straight axis depending on the presence of plane, parabolic, elliptic subsets in their boundary, and nonconvex object parts by generalized cones with curved axis. This approach allows to obtain a primitive based shape description after the decomposition of the object shape through the more general boundary-based methods. We present examples of decomposing and describing shapes of common objects in terms of their parts, subparts and associated features.

Knowing the end-effector location (position and orientation) and the joint angles of the robot manipulator in real-time will assist the manipulator in negotiating around the obstacles when the manipulator is moving in a crowded environment. Thus, Forward and Inverse Kinematics Computations (FKC and IKC) play very important roles in robotic manipulators. The main objective of this paper is to demonstrate the capability of learning different trajectories of the robot reachable space by using the proposed PNN model. A software package has been developed for solving both FKC and IKC. The software can discover both the structure and the coefficients of a model to describe the dependent output variables in terms of the independent input variables identified by the users. The simulation is performed in a two degree-of-freedom manipulator. The solutions of the built FKC and IKC networks are compared with the analytic equations. The PNN learns successfully the indicated path. The simulation result shows that the PNN can interpolate the indicated path better than 99.87% of accuracy by only training the built PNN network 361 data pairs (out of 2D space point). The approach presented here can be expanded to six degree-of-freedom type of manipulators. Detailed algorithms of the GMDH to construct the PNN kinematics models will be discussed.

We chose to tackle the problem of trying to set up time tables of the first 5 years of UTC curriculum, for several reasons: (1) the problem has eluded Operational Research combinatorial techniques for over 6 years, and Artificial Intelligence techniques for over 4 years, despite an average of 2 attempts per year, (2) there is a large amount of data (teaching personal curriculum structure, available classrooms,...), (3) there is no clear optimizing criterion (one must need a maximum flexibility for a student population which has not decided yet which courses it will take), (4) if constraints are taken literally, then the problem has no solution, (5) the actual manual solution involves a consensus between several decision makers, (6) an expert was available and willing to participate in the project. The number of students and of options increasing over the years, the problem is more and more difficult to visualize and to solve efficiently. Several techniques have been tested to solve this problem, but none succeeded. We first tried combinatorial techniques, but the problem has no solution if one cannot violate some of the constraints, and it was not possible to succeed. We tried to use Relational Data Bases to store the several data (more than 5000), but it was not efficient enough. Then, an Object Oriented Data Base gave us the good answer for storing all these data. We tried to use Prolog to solve the problem, but we realized that Backtracking and Backward inferences were not a good solution. We needed an expert system with a forward strategy. Thus, we implemented an expert system integrated into an object oriented data base. The software architecture was correct, but we forgot an important factor: what was the human expert's knowledge? How was it possible to 'catch' that knowledge? In fact, the knowledge base that we used was inefficient, and the system failed one more time. It has good performances during experimental tests. But in fact, it was not able to find a solution when running in real conditions, with a lot of data and constraints. Today, a new project is being implemented. We kept the last software architecture (object oriented data base, and expert system with a forward strategy), and also added to it Operational Research combinatorial techniques. We have been working for 2 years with the human expert, to improve the knowledge base, and the system is starting to give good answers, even in real conditions of work. This paper will describe this planning and scheduling problem, how we have been failing for 10 years, and how we are now starting to find a good solution, based on both Operational Research combinatorial and

We present current work on the Weather and Doctrinal Information Fusion (WADIF) system for Intelligence Preparation of the Battlefield (IPB). The task of the WADIF system is to determine the performance of an Army operational asset, such as a tank, taking into account current and projected terrain and weather conditions. This is a task that is currently performed manually, requiring significant personal resources. The WADIF system automates much of the process, enabling the user to analyze the area of interest in greater detail. By giving the user this ability, the WADIF system will significantly improve the quality of the decisions made. We present a discussion of the task environment and the architecture of the WADIF system in detail, and conclude by discussing future directions.

This paper describes an expert system that assists photolithography inspectors during IC fabrication process. Strict control of the process line is required, lack of control often results in costly errors. In photolithography, flaws that occur during the IC patterning process should be correctable within the same step. The diagnosis and trouble shooting of these flaws are done using the Photolith Analysis and Control (PAC). It is assumed that the readers have a good knowledge of IC processing and are familiar with the technical terminology prevalent in the field. The paper is arranged as follows, section 2 described the chief motivating factors for PAC. This is followed by a description of the user and video interfaces in section 3. Section 4 describes the PAC system in detail, section 5 contains performance and evaluation data and is followed by a conclusion.

The subject of control as it applies to a process or product embodies two components, namely detection and correction. The primary objective of this proposed interval/index is to combine the concepts of 'Shape and Location' in the form of an interval that will become a useful tool for control. This concept was developed to improve the often criticized indices that are in use today. Concerns associated with the standard Capability Index values, are based on the comments coming from the user population.

We present a knowledge based systems approach to the problem of the valuation of shares of private companies. At present the experts involved in this field perform their valuation manually and also treat each private company valuation as unique. Our approach provides the expert with several features: (1) The system is much faster than the current manual techniques used, with all calculations and matches taking place in under 10 seconds (on our current prototype). (2) The calculation and matching process is thorough and allows the expert to validate all results obtained. Also, due to the automated nature of the system there is not the possibility for steps in the calculation and matching process to be omitted; as is the case with the manual technique. (3) Due to the nature of the way that the system holds its information on the private companies, many companies data can be held at the same time and companies are selected for valuation when required. This feature lends itself to the larger accountancy firms which hold portfolios of clients. We have completed our initial prototype, based on a system design produced earlier by M. Jamurtas, and are currently in the process of having this validated and acceptance tested by experts out in the field.

The development of an intelligent mobile robot capable of assisting a certain group of the handicapped is presented in this paper. The system consists of three units: a robot, a computer, and a user. All three units communicate completely by remote, using RF signals. The user and robot communicate indirectly through the computer, using natural voice. Each user command is decomposed int a sequence of motion commands, using an object-oriented knowledge base. Robot motion is planned based upon inputs from ultrasonic sensors. The system is capable of learning an unknown, albeit somewhat restricted environment.

When a vision sensor is used to provide the sensory input of an autonomous vehicle, most researchers rely on a monocular camera system. In this paper, we have compared the capabilities and limitations of a monocular camera system with respect to a binocular camera system for the purpose of performing autonomous vehicle tracking. The problem of vehicle tracking includes automatic speed and steering control of an unmanned vehicle following the motion of a lead vehicle. We have indicted how the relative position of a lead vehicle is computed visually and investigated the quantization, tilt angle, pan angle, and road slope errors associated with a monocular and a binocular vision sensor. We have shown that a binocular camera system provides a more robust sensing mechanism when operating under realistic outdoor conditions.

This paper presents a new approach for a mobile robot that uses sonar and vision sensing to recognize indoor scenes. The complementary nature of these sensing modalities can provide complete information about the observed world, which is not available if either sensor is used alone. This approach is based on using some simple rules rather than mathematical models to integrate acoustical and optical sensory data for building a coherent representation of the robot's environment.

This paper presents an action in mission SOCRATES whose aim is the development of a self-calibration method for an autonomous mobile robot. The robot has to determine the precise location of the coordinate system shared by its sensors. Knowledge of this system is a sine qua non condition for efficient multisensor fusion and autonomous navigation in an unknown environment. But, as perceptions and motions are not accurate, this knowledge can only be achieved by multisensor fusion. The application described highlights this kind of problem. Multisensor fusion is used here especially in its symbolic aspect. Useful knowledge includes both numerous data coming from various sensors and suitable ways to process these data. A blackboard architecture has been chosen to manage useful information. Knowledge sources are called agents and the implement physical sensors (perceptors or actuators) as well as logical sensors (high level data processors). The problem to solve is self- calibration which includes the determination of the coordinate system R of the robot and the transformations necessary to convert data from sensor reference to R. The origin of R has been chosen to be O, the rotation center of the robot. As its genuine location may vary due to robot or ground characteristics, an experimental determination of O is attempted. A strategy for measuring distances in approximate positions is proposed. This strategy must take into account the fact that motions of the robot as well as perceptions may be inaccurate. Results obtained during experiments and future extensions of the system are discussed.

This paper presents a terrain model acquisition algorithm for a mobile robot with finite-range sensors in planar terrains. The finite two- dimensional terrain is populated by a finite number of stationary polygonal obstacles. If the robot can see only partial obstacle edge(s) because of its limited range of visibility, our algorithm guides the robot toward the direction in which the incomplete edge of the obstacle is completed at the end vertex. If the robot can see nothing at all within the current range, it is guided in a spiral-like manner to search the terrain for obstacles. In this paper, we formally describe this algorithm and show how it performs with respect to travel distance and the number of scanning operations for various sensor ranges. Also, some examples are shown.

We implement an architecture for the planning and control of an intelligent autonomous mobile robot which consists of concurrently running modules forming a hierarchy of control in which lower-level modules perform 'reflexive' tasks while higher-level modules perform tasks requiring greater processing of sensor data. A knowledge-based system performs the task planning and arbitration of lower-level behaviors. This system reasons about behavior selection (fusion) based on its current knowledge and the situation at hand provided by monitoring the status from lower-level behaviors and the map builder. We implement this knowledge-based planning module in CLIPS (C Language Implementation Production System), a rule-based expert systems shell. CLIPS is written in and fully integrated with the C language providing high probability and ease of integration with external systems. We discuss implementation issues including the implementation of control strategy in CLIPS rules and interfacing to other modules through the use of CLIPS user-defined external functions.

This paper describes a distributed architecture consisting of independent processed (agents) running in parallel and using assembly robotics as an experimental testbed. This architecture can support several kinds of organizational structures and of cooperation policies. An agent is described in an object-oriented environment and resides in a VAX processor connected to a local area network. We describe the kernel facilities replicated for all agents in order to participate in group problem solving and how they coordinate their efforts by exchanging information. We discuss in particular a heterarchical group organization where cooperation is made on a request basis.

A prototype solving the Shortest Path Problem (SPP) by a parallel message-passing algorithm is presented. The system, an OCCAM program running on a transputer board hosted by a PC, implements a known distributed algorithm for the SPP, based on the 'diffused computation' paradigm. A new parallel message-passing architecture is proposed, built upon a static packet-switching network with a fractal topology. The recursive, unlimited network, features an interesting property when applied to four-link processors (like transputers): it's decomposable, at any hierarchical level, in four-link modules or supernodes. Labelling and routing algorithms for the network, exploiting self-similarity, are described. Experimental results, obtained with a single transputer solving irregular random graphs (up to 256 nodes) are presented, showing a time complexity function growing linearly with the total number of arcs.

The lane sensing function detects and estimates the location of road and lane boundaries and is a core function of the Highway Driver's Assistant project within the Intelligent Vehicle Highway System program. The previously developed Hough Transform algorithm for lane sensing was improved significantly by embedding the knowledge of the road geometry and by reducing the area of the transform space. A modified Sobel edge- operator was also developed for detecting edges in the dark to light transition areas. The computational processing time was reduced from 30 seconds/image to 3 seconds/image on a VAX 8600 computer by using temporal knowledge of vehicle motion. The algorithm's improved performance was checked successfully on over 800 images of highway scenes.

Methods based on various morphological image processing techniques are developed for the discrimination of images reflected by road surfaces in different states. Numerous texture analysis methods are also applied tot he texture classification of the same images. Some of the methods developed are shown to be very efficient for automatic identification of the road surface states, whatever the granular structure of the road surface may be.

The handwritten signature is the most widely employed source of secure identification in the United States, especially for cashing checks, and verifying credit card transactions. Currently, all signature verification is based on visual inspection by a teller or a store clerk. Previous successful techniques for forgery detection have primarily been on-line techniques. This research is an extension of the first author's work on forgery detection and describes an algorithm to detect forgeries perpetrated by using a tracing paper or a glass plate. This algorithm is very successful when used in conjunction with the algorithm developed earlier by the first author.

A method to find the so-called missing dots in a heliotest strip has been developed. The developed method bases on image processing and supervised learning. The performance of the method is good or satisfactory in 85 cases of hundred. The equipment needed consist of a 386-based MS/DOS-computer, commercial image processing board and software, commer cial coordinate table, CCD-camera and lighting equipment.

For the task of knowledge acquisition in mechanical engineering redesign a computer-aided system is described. It elicits design knowledge direct from the designer by evaluating the design protocol and comparing the result with existing solutions to similar problems. The solutions are represented by means of spatial relations between the components of a part which allow for the definition of similarity between different constructions.

The aim of this work is to build a machine learning system. A new learning method is proposed. The complete system contains a Task Planner, a Problem Generator, a Plan Scheduler and an Apprentice. This system solves problems and at the same time learns to improve its behavior from the experience. The task planner is based on a production system. It uses a parallel planning method as a forward search strategy by matching the domain rules for all the participants of the model. When the planner finds a problem state where it is possible to apply one operator to several participants of the world, it decides to execute only one operation. Then, the other parts of the complete system start to work. The Problem Generator and the Plan Scheduler obtain all the solutions plans associated to the current problem. When the learning system knows which is the best plan it searches all the plans associated with the best plan, we mean, all the possible solutions of the problem starting from the same state of the world. The proposed learning method looks for differences between the steps of these plans and makes up the goal concept. Then, it explains this goal concept finding out differences or similarities between the value of the attributes of the participants implied into this world state.

We consider verifying knowledge bases to three levels of rigor: detection of anomalies, verification of safety properties, and verification of full correctness. We present formal definitions for four classes of anomalies which may be present in knowledge bases expressed using first order logic: redundancy, ambivalence, circularity and deficiency. The definitions are initially given for rule-based systems without uncertainty, but we extend them to consider uncertainty and frame-based knowledge representations. We demonstrate that, although verification of full correctness will not usually be feasible for knowledge-based systems, it is important that their safety properties be verified, and we present a method for doing this based on our definitions of logical anomalies. We demonstrate the validity of this framework by presenting the results of a verification performed on the knowledge base of a working expert system.

In this paper, we propose an internal protocol assistant that can help users to verify and validate the cooperation and the relationships of the messages exchanged among agents in a distributed office information system to see if internally the underlying agent can work properly or not when some specific events occur. Office Procedure Model (OPM), which is the formal model of the concerned agent in an office information system, is used as an example. A connection matrix is constructed to represent the OPM diagram. The algorithm for partitioning the OPM diagram into several message groups to reduce its scope (called Indivisible Cut Zones) is also introduced. Finally, for each message group we set up a Transition Table from which the expressions of the objects in ,the diagram can be easily derived. Users can inspect the expressions one by one to verify and validate their systems.

This paper presents a knowledge engineering shell, KEshell. It adopts a 'rule skeleton + rule body' representation and a linear forward reasoning algorithm. The system structure, the representation scheme, the inference engine and the interactive knowledge acquisition procedure are described. Some further attempts are also briefly outlined.

Range images can provide very useful information for operating robotic systems in cluttered and dynamic environments. Development of an integrated mobile robotic system which can acquire, process, analyze and perform actions based upon such analysis, in real time, involves the resolution of many complex issues. In this paper we describe research activities to demonstrate development and operation of an autonomous mobile robotic system capable of detecting 55 gallon drums in range images, and then performing complex radiation survey tasks on the detected drums. The robot must perform tasks such as precision navigation with crabbing, object identification and localization, docking with the located object, and scanning of the object for radiation. The processing of range data involved five steps: (1) correction of the cosine effect, (2) image filtering, (3) calculation of surface normals, (4) image segmentation, and (5) object recognition and localization. The results of the range processing procedures are presented. The novelty of the techniques is that they are being adapted, combined, and applied to real data obtained form a sensor-based mobile robot in a cluttered environment. Experimental results demonstrating the robustness of the approach on real range data are presented. We also present some experimental results of a Geometric Modeling and Sensor Simulation System for generating synthetic range imagery using 3D world models.

This research describes a robust, efficient, and real-time computer vision system that can automatically inspect defects of protruded print characters on injection molded plastic, typewriter print wheels. Possible defect types include insufficient fill, voids, and cracks. These defects can be described as poor edge sharpness, large edge position deviation from an established standard, and irregularities of the inside surface. Template matching is used for character detection and extraction. Matching performance measurements are used to evaluate closeness with respect to a reference print character of accepted quality. A hierarchical structure is used to improve the robustness of position detection and acceptable performance measures in knowledge rules are incorporated to increase the speed of the search. Characters are extracted from the image by a logical 'AND' operation in which a filled, slightly enlarged, uniform gray scale pattern of the print character is used as a template. Feature extraction and matching is done by using a distance image template matching technique which makes the system more robust and effective. Finally, a set of matching measurements is extracted to determine edge sharpness, edge deviation, and smoothness of the inside surface and local matching measurements are used for determining the detail of defects.

The aspect graph representation of different classes of objects has been a popular research topic in the last few years. Algorithms to compute the aspect graph have been developed for polyhedra, curved-surface objects, and articulated assemblies of rigid parts. This paper discussed the problem of how to compute the aspect graph for nonrigid or deformable objects.

In this paper, an object segmentation algorithm is presented to extract partially occluded objects from a scene. The visual scene is composed of a group of 3D overlapping objects which can be represented by straight- line drawings projected on 2D image plane. The goal is to decompose the visual scene into several 3D objects by extracting line drawing from partially occluded scene. The criteria used for segmenting a visual scene into composing objects include junction type interpretation and region homogeneity. The segmented result can be used as a basis for partial matching in recognizing partially occluded objects.

This paper presents a new approach to integrated modeling of three dimensional objects for generating visual and thermal images under different viewing, ambient, and internal conditions. A volume surface octree is used for object modeling. It is shown that the above representation is suitable for thermal modeling of complex objects with non-homogeneities and heat generation. The technique to incorporate non- homogeneities and heat generation using octree intersection is described. The proposed model may be used to predict a discriminatory feature for object recognition based on the surface temperature and intrinsic thermal properties in any desired ambient condition. The model is designed to be used in a multi-sensor vision system using a hypothesize and verify approach. Several examples of the generated thermal and visual images are presented, which illustrate the usefulness of the approach.

This paper describes an algorithm which automatically creates an accurate and realistic reconstruction of buildings from high resolution, large-format digitized aerial stereo photographs. The system makes use of both reflectance and disparity data. Given a stereo pair of reflectance images of a scene containing buildings, the program automatically builds up a wireframe description of buildings and outputs a modified elevation map for the scene in which building edges are reconstructed accurately. A wide spectrum of computer vision techniques have been employed in this system- a fast stereo correlation technique; a robust, discontinuity preserving surface approximation algorithm to patch in uncorrelated areas; and knowledge based vision techniques to segment buildings from background and then reconstruct the building. A geometric model, in the form of a wing-edged data structure, is used to keep the wireframe while building up the structure. We also employ texture mapping to overlay the reflectance image back on the elevation data in order to provide a realistic display.

We describe a new method for shape decomposition which relies exclusively on global properties of the surface which are fully characterized by local surface properties. We propose that a useful parcellation of shapes into parts can be obtained by decomposing the shape boundary into the largest convex surface patches (LCP) and the smallest nonconvex surface patches. The essential computational steps of this method are the following (i) build initial parts from the largest locally convex patches, (ii) consider an initial part as a constituent pat if it is essentially convex, and (iii) obtain the remaining constituent parts by merging adjacent initial parts generated by the largest locally convex and the smallest convex patches of nearly the same sizes. We show that the decomposition of shapes into the largest convex patchesaims to maximize the 'thingness' in an object, and to minimize its 'non-thingness', that is the method is conductive to a natural parcellation of shapes into constituent parts useful for recognition and for inferring function.

This paper presents a methodology and algorithms for visual reconstruction from the 2D string representation of an image. The 2D string representation consists of symbolic objects and spatial operators representing the spatial relations among objects or subobjects in an image. The method proposed here is to reconstruct the symbolic picture using visual reasoning. First, 2D strings, spatial operators, and projection rules are presented. Then we present a visual reasoning approach with which the algorithm for visualization is developed. The rules for visual reconstruction are then described in detail. Finally, we describe a prototype visual reasoning system which reconstructs images from 2D strings. We also present some promising experimental results and discuss applications of this approach.

This paper describes two new parallel algorithms for thinning. Instead of a conventional 3 X 3 operator, a small cross window '+' with 5 or 9 points is employed in the implementation. The algorithms do not change the connectivity of the images. The experimental results show that the algorithms make the skeleton much closer to the medial axis, and make it more convenient to reconstruct the original pattern. Comparing with the other thinning methods, the new approach here gives advantages on saving the time and the window size when real world images are processed.

Correction-based template matching has been used extensively in computer vision for object recognition and also for other tasks such as edge detection, stereo, motion and inspection. It has also found wide application in character recognition. A deeper understanding of the performance of this technique for such tasks would help predict when it will succeed or fail. Previous work on this problem has examined correlation-based template matching using signal processing techniques. Our approach is different: we dissect it employing concepts from geometry and physics. This leads to new insights into correlation-based template matching. We study the performance of correlation between images for different lighting conditions, viewpoints and scales of a scene, obtaining new results for scale variation and viewpoint change for binary images. We analyze gray level images for changes in lighting alone and obtain useful and novel formulae. Knowing how correlation behaves with these changes helps to strategically distribute templates for a given recognition task. We then develop a method to compute the probability of confusion for recognition by template matching. We obtain a closed form solution for the probability of confusion in the two template case. We conclude by noting that template matching encounters difficulties in tasks such as object recognition because of its strong dependence on viewing conditions, although it can be useful in some situations when templates are chosen and positioned judiciously.

This paper describes a position and rotation invariant fast object classification scheme. A parallel region growing technique is used to detect objects in binary images. 2D fast Fourier transform (FFT) is applied to each object region after translating the origin of the image coordinate system to the object center and aligning the image coordinate axes with the object's principal axes. The first five components from the principal lobe of the Fourier spectrum of each object are selected as characteristic features for minimum-distance classification. For time efficiency, region growing and 2D FFT computations were performed on a 16-node hypercube processor.

This paper describes the use of Hadamard transform for recognizing objects in an industrial environment. The transform is implemented on a 16-node hypercube array processor using ring topology. An adaptive algorithm extracts the Hadamard domain features that best represent a particular object independent of its size, position, and orientation. The first 8-10 Hadamard coefficients are selected for object classification using the minimum-distance rule. A tenfold speedup is achieved by parallelizing the algorithms for extracting the Hadamard domain features.

A stochastic neural net shares with the normally defined neural nets the concept that information is processed by a system consisting of a set of nodes (neurons) connected by weighted links (axons). The normal neural net takes in inputs on an initial layer of neurons which fire appropriately; a neuron of the next layer fires depending on the sum of weights of the axons leading to it from fired neurons of the first layer. The stochastic neural net differs in that the neurons are more complex and that the vision activity is a dynamic process. The first layer (viewing layer) of neurons fires stochastically based on the average brightness of the area it sees and then has a refractory period. The viewing layer looks at the image for several clock cycles. The effect is like those photo sensitive sunglasses that darken in bright light. The neurons over the bright areas are most likely in a refractory period (and this can't fire) and the neurons over the dark areas are not. Now if we move the sensing layer with respect to the image so that a portion of the neurons formerly over the dark are now over the bright, they will likely all fire on that first cycle. Thus, on that cycle, one would see a flash from that portion significantly stronger than surrounding regions. Movement the other direction would produce a patch that is darker, but this effect is not as noticeable. These effects are collected in a collection layer. This paper will discuss the use of the stochastic neural net for edge detection and segmentation of some simple images.

Two different approaches have been implemented to model the filling-in procedure for inferring the shape of occluded objects: computational and back-propagation network. The two approaches have been found to have functional similarities. The receptive fields of the units have been analyzed and an analogy with the receptive field of neurons in area V2 of visual cortex was suggested.

Neural network technology is finding applications in a wide range of research fields, such as, pattern recognition, robot navigation, communications, computer vision, etc. Neural nets can also be used as experts in a particular problem domain. Powerful learning algorithms associated with neural net architectures provide them the ability to extract similarities from the database and encode these properties in a weight matrix. This reduces the dependance on a human expert to create a rule base. In this paper we describe a knowledge-based network for the diagnosis of the risk factor of developing coronary atherosclerosis. The entire system consists of a knowledge based system which uses a neural network, that is trained for a specific set of data, to obtain a risk factor which it then modifies based on additional information to obtain the final result.

Expert systems and other conventional approaches have proven to be of limited ability in addressing the problem of robot navigation. Recent advances in neural network technology, in particular, powerful learning paradigms and neuro-computer hardware, could provide crucial tools for developing improved algorithms and computational hardware for robot navigation. Several researchers have designed nonlinear controllers using neural networks for precise navigation and positioning of the robotic vehicles around fixed and moving objects. This paper reviews various neural net applications in the areas of robot control. Autonomous land vehicles, and underwater robotic vehicles. The results show the feasibility of using neuro-controllers for these robotic vehicles in the presence of unpredictable changes int eh dynamics of the vehicle and its environment.

Physical models for color image formation provide constraints which are useful for interpreting 3D scenes. I summarize the physics underlying color image formation. Models for surface and body reflection from metals and dielectrics are analyzed in detail. This analysis allows us to evaluate the benefits we stand to gain by using color information in machine vision. I show from the reflection models that color allows the computation of image statistics which are independent of scene geometry. This principle has been used to develop an efficient algorithm for segmenting images of 3D scenes using normalized color. The algorithm applies to images of a wide range of materials and surface textures and is useful for a wide variety of machine vision tasks including 3D recognition and 3D inspection. Experimental results are presented to demonstrate the scope of the models and the capabilities of the segmentation algorithm.

In this paper, we present a knowledge-based texture segmentation system for the identification of 3D structures in biomedical images. The segmentation is guided by in Iterative Octree Expansion and (leaf node) Linking control algorithm. The segmentation is performed directly in 3D space which is contrary to that done in common approaches where 3D structures are reconstructed from results of 2D segmentation of a sequence of consecutive, cross-sectional images. Test result of a prototype of this system on real data confocal scanning fluorescence microscopic images of a developing chick embryo heart is reported.

Based on a two-view vision system technique and basic geometrical principles, a new algorithm for 3D texture recovery is proposed in this paper. The basic principle is to acquire two projected images from a surface and derive the characteristic line vector from each projection followed by the estimation of the normal vector of the surface. Also a new method for the estimation of the characteristic line vector is presented, called the fanshape pyramid structure algorithm. The algorithm has proved to superior to the traditional spectrum method. It can easily be extended to the multiview case or to the processing of image sequences in order to yield an improved estimation. In addition, the algorithm offers reduced computation and a simple structure for parallel implementation.

The prototypic nature of knowledge base systems has made it difficult to realize a 'production environment knowledge base system', until now. We will describe the use of computer vision and knowledge base system (KBS) techniques to delineate, identify, and attribute cartographic features from digital stereo imagery, with the distinction that these functions are implemented in a large production environment. Along with a discussion of the technology that supports this application, we broach the realities of using KBS technology in large production environments. The KBS chiefly assists in the cartographic feature identification and attribution. Our paper will cover the knowledge acquisition activities and the selection knowledge representations that supported the construction of the knowledge base as well as the inference engine that supports the operation of the KBS. We will describe in detail the computer vision tools and the level of image understanding that is achieved by their application. This includes a discussion of the 'computer vision tool box' used for the delineation of cartographic features. Of particular interest are the variety of technologies that support the tool box, such as the use of artificial neural networks. Because of their significance to the real world success of KBS technology, we include the subjects of risk mitigation in the design phase in addition to the ongoing KBS support in the maintenance phase.

This paper describes a computer-oriented representation for sentence information. Whereas many Artificial Intelligence (AI) natural language systems start with a syntactic parse of a sentence into the linguist's components: noun, verb, adjective, preposition, etc., we argue that it is better to parse the input sentence into 'meaning' components: attribute, attribute value, object class, object instance, and relation. AI systems need a representation that will allow rapid storage and retrieval of information and convenient reasoning with that information. The attribute-of-object representation has proven useful for handling information in relational databases (which are well known for their efficiency in storage and retrieval) and for reasoning in knowledge- based systems. On the other hand, the linguist's syntactic representation of the works in sentences has not been shown to be useful for information handling and reasoning. We think it is an unnecessary and misleading intermediate form. Our sentence representation is semantic based in terms of attribute, attribute value, object class, object instance, and relation. Every sentence is segmented into one or more components with the form: 'attribute' of 'object' 'relation' 'attribute value'. Using only one format for all information gives the system simplicity and good performance as a RISC architecture does for hardware. The attribute-of-object representation is not new; it is used extensively in relational databases and knowledge-based systems. However, we will show that it can be used as a meaning representation for natural language sentences with minor extensions. In this paper we describe how a computer system can parse English sentences into this representation and generate English sentences from this representation. Much of this has been tested with computer implementation.

LCS (Language Comprehension System) is a software package designed to improve man-machine communication with computer programs. Different simple structures and functions are available to build man-machine interfaces in natural language. A user may write a sentence in good English or in telegraphical style. The system used pattern matching techniques to detect misspelled words (or badly typed words) and to correct them. Several methods of analysis are available at any level (lexical, syntactic, semantic...). A special knowledge acquisition system is used to introduce new works by giving a description in natural language. A semantic network is extended to a representation close to a connexionist graph, for a better understanding of polysemic words and ambiguities. An application is currently used for a man-machine interface of an expert system in computer-aided education, for a better dialogue with the user during the explanation of reasoning phase. The object of this paper is to present the LCS system, especially at the lexical level, the knowledge representation and acquisition level, and the semantic level (for pronoun references and ambiguity).

Most connectionist parsers either cannot guarantee the correctness of their derivations or have to simulate a serial flow of control. In the first case, users have to restrict the tasks (e.g. parse less complex or shorter sentences) of the parser or they need to believe in the soundness of the result. In the second case, the resulting network has lost most of its attractivity because seriality needs to be hard-coded into the structure of the net. We here present a hybrid symbolic connectionist parser, which was designed to fulfill the following goals: (1) parsing of sentences without length restriction, (2) soundness and completeness for any context-free grammar, and (3) learning the applicability of parsing rules with a neural network. Our hybrid architecture consists of a serial parsing algorithm and a trainable net. BrainC (Backtracking and Backpropagation in C) combines the well known shift-reduce parsing technique with backtracking with a backpropagation network to learn and represent the typical properties of the trained natural language grammars. The system has been implemented as a subsystem of the Rochester Connectionist Simulator (RCS) on SUN- Workstations and was tested with several grammars for English and German. We discuss how BrainC reached its design goals and what results we observed.

A new algorithm is presented to extract the chain-code representation of the boundary of a binary image. The quadtree representation of this image acts as the input to the algorithm. This algorithm is more efficient than the previously reported algorithm. The algorithm can trace the boundary according to either eight-connected or four-connected neighborhood definitions without any modification. The boundary is traced in the clock-wise direction with respect to the object. The algorithm is based on the following two observations; (1) Visiting a sequence of four-connected white nodes that lie along the object boundary yields an eight-connected boundary while (2) visiting a sequence of four-connected black nodes that lie along the object boundary yields a four-connected boundary. In other words, tracing the object boundary from outside the object using four-connected neighborhood definition extracts an eight-connected boundary, and tracing the object boundary from within the object using four-connected neighborhood definition extracts a four-connected boundary.

This paper presents an integrated image segmentation method using edge and needle map which compensates deficiencies of using either edge-based approach or region-based approach. Segmentation of an image is the first and most difficult step toward symbolic transformation of a raw image, which is essential in image understanding. In industrial applications, the task is further complicated by the ubiquitous presence of specularity in most industrial parts. Three images taken from three different illumination directions were used to separate specular and Lambertian components in the images. Needle map is generated from Lambertian component images using photometric stereo technique. In one channel, edges are extracted and linked from the averaged Lambertian images providing one source of segmentation. The other channel, Gaussian curvature and mean curvature values are estimated at each pixel from least square local surface fit of needle map. Labeled surface type image is then generated using the signs of Gaussian and mean curvatures, where one of ten surface types is assigned to each pixel. Connected regions of identical surface type pixels provide the first level grouping, a rough initial segmentation. Edge information and initial segmentation of surface type are fed to an integration module which interprets the edges and regions in a consistent way. During interpretation regions are merged or split, edges are discarded or generated depending upon global surface fit error and consistency with neighboring regions. The output of integrated segmentation is an explicit description of surface type and contours of each region which facilitates recognition, localization and attitude determination of objects in the image.

The problem of model-based object recognition is considered as a computational process incorporating a means of clustering feature data consistent with the parts of a structural shape model. A general approach is developed that using both continuity and shape constraints for fitting axial-curve models to derived feature patterns. This integrated approach allows for noise datums to be disregarded, while missing data can be inferred by the interpretation of axial point sequences. Complete object structures are recovered using a circular operator to detect features of shape discontinuity (corners, junctions and tips). The approach is demonstrated on images from various domains, with the main result being a suburban road network analysis of a high resolution aerial image. Other results include overlapping circles, circuit board traces, and a LANDSAT image of the Mississippi river.

Using polynomial to represent binary images, we have developed algorithms to perform a number of standard operations. It is shown that all the morphological operations on images can be done using the polynomial approach. Further, we extend our methods to process 3D objects.

We write about the relationship between numerical patten recognition and neural-like computation networks. Extensive research that proposes the use of neural models for a wide variety of applications has been conducted in the past few years. Sometimes justification for investigating the potential of neural nets (NNs) is obvious. On the other hand, current enthusiasm for this approach has also led to the use of neural models when the apparent rationale for their use has been justified by what is best described as 'feeding frenzy'. In this latter instance there is at times concomitant lack of concern about many 'side issues' connected with algorithms (e.g., complexity, convergence, stability, robustness and performance validation) that need attention before any computational model becomes part of an operation system. These issues are examined with a view towards guessing how best to integrate and exploit the promise of the neural approach with there efforts aimed at advancing the art and science of pattern recognition and its applications in fielded systems in the next decade.

A parallel architecture for rule-based systems is generally based on techniques that reduce the execution time of a production cycle, match- select-act cycle. With these techniques, speedups in rule-based systems are limited to tenfold. We investigate a new approach to achieve higher speedups by developing a computer architecture that exploits parallelism at much higher levels. This high-level parallel architecture is based on a semantic network representation and message passing. This architecture closes the semantic-gap that exists between the application and its implementation and yields higher speedups. To demonstrate our approach, we have chosen the VLSI architecture simulation application which has inherent parallelism and demands large execution times on conventional computers. We performed functional simulations for two applications and measured close to linear speedups. In this paper, we present the applications, knowledge representation scheme, architecture suitable for this set of applications, and speedup measurements.

In this paper, we present an execution scheme allowing a direct and a pipeline evaluation of a Prolog Program. The execution scheme enhances Prolog performances in interpreted mode, by means of associative processing tools embodied in Content Addressable Memories and potential parallelism existing between clauses selection, unification, and access to clause arguments. The interpretation algorithm is distributed on several processing units, which are Content Addressable Memories (CAMs). These latter are generic and reconfigurable dealing with much more Artificial Intelligence applications, through improved target languages like Prolog, Lisp, and Object oriented languages. The model has been evaluated with a functional simulator written in Le-lisp. The results show the CAMs feasibility in improving Prolog execution at performances greater than 160 KLIPS, in interpreted mode.

This report discusses ongoing work on a parallel reduced-instruction- set-computer (RISC) architecture for automatic production matching. The PRIOPS compiler takes advantage of the memoryless character of automatic processing by translating a program's collection of automatic production tests into an equivalent combinational circuit-a digital circuit without memory, whose outputs are immediate functions of its inputs. The circuit provides a highly parallel, fine-grain model of automatic matching. The compiler then maps the combinational circuit onto RISC hardware. The heart of the processor is an array of comparators capable of testing production conditions in parallel, Each comparator attaches to private memory that contains virtual circuit nodes-records of the current state of nodes and busses in the combinational circuit. All comparator memories hold identical information, allowing simultaneous update for a single changing circuit node and simultaneous retrieval of different circuit nodes by different comparators. Along with the comparator-based logic unit is a sequencer that determines the current combination of production-derived comparisons to try, based on the combined success and failure of previous combinations of comparisons. The memoryless nature of automatic matching allows the compiler to designate invariant memory addresses for virtual circuit nodes, and to generate the most effective sequences of comparison test combinations. The result is maximal utilization of parallel hardware, indicating speed increases and scalability beyond that found for course-grain, multiprocessor approaches to concurrent Rete matching. Future work will consider application of this RISC architecture to the standard (controlled) Rete algorithm, where search through memory dominates portions of matching.

Providing robot hands with the intelligence required for effective multifinger grasping is a difficult problem. In particular, a good choice of grasp points on an object to be manipulated is critical in order to achieve dexterous manipulation. A methodology and algorithmic implementation is proposed for the choice of the feasible grasp points on irregular objects. The main thrust of this analysis is the embedding of grasp mechanics - namely intuitive human-like force distribution and computational efficiency - into a practical methodology for intelligent grasp planning. The strategy of grasp reasoning is to plan the grasp so that the fingers will be guaranteed to firmly grasp the object during handling and generate fine motion to perform tasks using the same planned grasp points. We also require the ability to change some of the grasp positions while holding the object firmly. The key to our approach is that the reasoning is based on a solid mathematical model of grasp mechanics - information from the mechanics of differing grasp candidates is automatically included at the grasp strategy stage. Therefore, our grasp has the advantage of being intelligence-based without sacrificing the physics of the grasp. Additionally, our underlying method for grasp mechanics is designed to be simple, computationally efficient, and intuitively natural, and can be easily employed in real time. Thus the grasp chosen is one that feels right, and is also based on solid physical principles, combining the intelligence of the expert with the mechanics needed for precise control.

In this paper we are focusing on the discrimination and recognition of arm movements described in terms of the evolution of the joint angles during the trajectory of the movement. Considering this as a classification problem, we could use a three layered feedforward network trained by the backpropagation learning algorithm to discriminate between arm movements described in the representation proposed by Marr and Vaina (1982) and Vaina and Bennour (1985). We discuss the contribution of the sampling rate for the value of the joint angles in the input layers, the number of angles necessary for a good recognition of the movement. We introduce a new method for determining the size of the network at input stage, which eliminates the unnecessary data in the training set. The effect of changing the number of frames used to train the network for each movement and the properties of the backpropagation algorithm in this application are discussed. We demonstrate that a description of movements at multiple scales of resolution organized from general to particular is conducive to an efficient learning scheme in the network. The main result obtained is that the discrimination between different arm movements is most efficiently obtained at the appropriate and most meaningful scale of resolution, as predicted by the Marr and Vaina's model.

Modern military surveillance systems typically include a number of different, independently adjustable sensors distributed throughout an environment to be monitored. These sensors should be configured so that their integrated outputs provide the optimal combination of probability of target detection and probability of false alarm. While it is desirable to optimize this measure of system performance, it is also desirable to minimize the enemy's ability to detect these sensors. These are conflicting goals. Each sensor can typically monitor only a small part of the environment and can sample only a small number of target discriminants. Because there are only a limited number of sensors available, sensor placement and configuration are critical to system performance. A system may use passive sensors to cue active sensors, or use low-resolution sensors to cue high-resolution sensors. All available information (properties of the sensors, properties of the environment being monitored, and known target locations and properties) should be used to determine an optimal sensor configuration. We call this the sensor cueing problem. This paper describes an algorithm that uses a heuristic search to efficiently solve the sensor cueing problem. The algorithm assumes that sensor locations are fixed in advance, but that other attributes (pointing direction, field of view, focus, etc.) may be adjusted to maximize system performance. Expected system performance is based on how well the group of sensors covers regions of the environment known to contain targets, as well as regions of the environment where targets are expected to appear. The algorithm's performance and possible extensions are described.

The task of path planning in mapped environments is well-known in robotics. Given an object and a set of obstacles located in space, the problem is to find a shortest path for the object from the initial position to the goal position which avoids collisions with obstacles along the way. In this paper, we will present a path planning algorithm to solve the find-path problem. The advantages of this algorithm are that it is quite fast and it finds the shortest path which avoids obstacles generously rather than avoiding them barely.

In this paper, two new methods of analyzing repertory grids are explored. The property of transitivity is used to help unravel latent relationships within a grid. Relationships are exposed with the help of transitive closures and principal component analysis. One method produces 'sensitivity maps' by superimposing principal component scatterplots of transitive closures of the (transformed) element (or construct) grid computed with six T-norms onto the principal component display of the original. The other method uses the well known (alpha) cut method to decompose each transitive closure into a partition tree of clusters of elements (or constructs). Both method are illustrated with a small numerical example.

There is a longstanding tradition in science of according much more respect to theories which are quantitative, formal and precise than to those which are qualitiative, informal and approximate in nature. In recent years, however, the validity of this tradition has been called into question by the emergence of artificial intelligence as one of the fundamentally important areas of modem science and technology. More specifically, what has become increasingly obvious is that many of the basic problems relating to the conception and design of complex knowledge-based and robotic systems do not lend themselves to precise solution within the framework of classical logic and probability theory. Thus, to be able to deal with such problems we frequently have no choice but to accept solutions which are suboptimal and inexact. Furthermore, even when precise solutions can be obtained, their cost is generally much higher than that of solutions which are imprecise in nature and yet yield results which fall within the range of acceptability. Seen against this background, fuzzy logic may be viewed as an attempt at formalization of approximate reasoning, which is characteristic of the way in which humans reason in an environment of uncertainty and imprecision. In this perspective, fuzzy logic may be viewed as a generalization of both multivalued logic and probability theory. In relation to these theories, its principal constituents are: (a) a meaning-representation system referred to as test-score semantics for representing the meaning of complex facts, rules and commands expressed in a natural language; and (b) an inferential system for inference under uncertainty which is applicable to knowledge-bases that are imprecise, incomplete or lacking in reliability. In contrast to classical logical systems, the inference processes in fuzzy logic are computational rather than symbolic. Thus, in general, inference in fuzzy logic reduces to the solution of a nonlinear program. This reflects the fact that in fuzzy logic a proposition is interpreted as a constraint on a variable, with constraint propagation playing the role of chaining and aggregation. A branch of fuzzy logic which plays a particularly important role in the representation and inference from commonsense knowledge is that of dispositional logic. As its name implies, this logic deals with dispositions, that is, with propositions which are preponderantly but not necessarily always true, e.g. ,birds can fly, seat belts work, Swedes are blond, etc. A related concept is thatof a subdisposition, e.g. , overeating causes obesity, which may be interpreted as an assertion concerning the increase in a conditional probability which is implicit in the defining proposition. At this juncture, most of the practical applications of fuzzy logic involve three basic concepts: (a) the concept of a linguistic variable, that is, a variable whose values are words or sentences in a natural or synthetic language; (b) the concept of a canonical form, which expresses a proposition as an elastic constraint on a focal variable; and (c) the concept of interpolative reasoning, which provides a means of filling in the gaps in knowledge from which an answer to a query is to be derived. These concepts serve as a basis for the description of qualitative dependencies between two or more variables through a system of fuzzy if-then rules in which the antecedents and consequents are expressed in a canonical form involving linguistic variables. The role played by fuzzy if-then rules in the applications of fuzzy logic is explained and illustrated by examples. Recent advances in fuzzy logic are analyzed and their potential applications are discussed.