Abstract: A first information transforming unit, capable of learning, transforms state information acquired from an object into a first intermediate representation. A second information transforming unit, capable of learning, transforms knowledge information derived, from attributes of the object, by a human being into a second intermediate representation. A fusing unit produces a common intermediate representation using the first intermediate representation and the second intermediate representation. The common intermediate representation is used as a teacher signal in order to train the first and second information transforming units. Thereafter, an intermediate representation made by transforming state information acquired from a certain object will be consistent with an intermediate representation made by transforming knowledge information derived from the object by a human being.

Abstract: The present invention provides a system for learning from and responding to regularly arriving information at once by quickly combining prior information with concurrent trial information to produce useful learned information. At the beginning of each time trial a vector of measurement values and a vector of measurement plausibility values are supplied to the system, and a learning weight is either supplied to or generated by the system.

Abstract: An authoring system and method create a knowledge base for an expert system that comprises communication cells for storing external data, assertion cells for storing data acquired from a user in the course of a user dialog, and interaction cells that specify (multimedia) interactions with the user, which are carried out as the cells are invoked by the run-time program. Each interaction cell has an enabling condition that must be satisfied before it can be invoked. At authoring time, the knowledge base is constructed by an expert system author who allocates the knowledge base cells on a rectangular grid similar to a spreadsheet. The author can try out various what-if scenarios by setting the values of assertion and communication cells and recalculating the knowledge base or simulating the created expert system. The grid is only an authoring tool, which is not visible or present at run time.

Abstract: A method and device for clock-synchronized motion control in discrete-time sampling systems is provided, in particular for use in numerical control systems for machine tools and robots. The method includes the derivation of any order m from the tool path feedrate (v or x.sub.2) or of any order m+1 from the position (s or x.sub.1) serving as the manipulated variable (u), the acceleration (a or x.sub.3) or the jerk limitation (r or x.sub.4) preferably being used. The exact control sequence (u) for the motion control is determined from the n=m+1 discrete-time system matrix equations of the sampling system, given an initial state (x.sub.k) and desired state vector (xx.sub.k+v) The manipulated variable (u) is changed in accordance with the specific values of the control sequence (u) within the time reference (k*T) of the sampling system, in particular in the interpolation clock cycle when working with numerical control systems for machine tools and robots.

Abstract: The invention relates to an apparatus for detecting fraud using a neural network. The architecture of the system involves first employing a conceptual clustering technique to generate a collection of classes from historical data. Neural networks are provided for each class created by the clustering step and the networks are trained using the same historical data. This apparatus is particularly useful for detecting the incidence of fraudulent activity from very large amounts of data such as tax returns or insurance claims.

Abstract: Apparatus and methods for machine learning the hypotheses used in the classifier component of pattern classification devices such as OCRs, other image analysis systems, and and text retrieval systems. The apparatus and methods employ machine learning techniques for generating weak hypotheses from a set of examples of the patterns to be recognized and then evaluate the resulting hypothesis against example patterns. The results of the evaluation are used to increase the probability that the examples used to generate the next weak hypothesis are ones which the previous weak hypothesis did not correctly classify. The results of the evaluation are also used to give a weight to each weak hypothesis. A strong hypothesis is then made by combining the weak hypotheses according to their weights.

Abstract: A system for computing an affinity for relocating a cell on a surface of a semiconductor chip is disclosed herein. The cell is located within a region and belongs to a net of cells. The system initially computes a weight associated with all cells in the net. The sytem then sums the weights of all cells in the net containing the cell for all cells located inside the region and at positions greater than and less than edges of the region and computes the affinity for moving the cell to points on the surface greater than, equal to, and less than the current position of the cell based on the weight sums from said summing function. The computing function further comprises combining the affinities determined based on weight sums with other affinities. The summing function further comprises computing a relationship between the amount of rows and columns of regions on the semiconductor chip surface, and the affinity computation function comprises combining the relationship with the weight sums.

Abstract: A knowledge processing system includes a memory for storing knowledge in a plurality of knowledge modules in a knowledge bank, in which the knowledge includes data associated with pre-existing solutions to previous problems; and a processing unit, operating knowledge-based software, and including a knowledge processing unit for generating a plurality of artificial intelligence based modules, and for executing the plurality of artificial intelligence base modules to process input data, including problem related data, for generating a plurality of solution-related modules from the data representing the pre-existing solutions, with the solution-related modules corresponding to solutions to the current problem; a numerical processing unit for generating a confidence level associated with a respective solution to the current problem; and a database processing unit for controlling the storing of the solution-related modules in the memory.

Abstract: This invention relates to a fuzzy processor having an input X for at least a plurality of input variables X-i and an output U for one or more output results U-k, and including a fuzzyfication unit FU having an input coupled to the input X, a fuzzy rule processing unit RU having an input coupled to the output of the fuzzyfication unit FU, and a defuzzyfication unit DU having an input coupled to the output of the processing unit CU and an output coupled to said output U, wherein the output of the defuzzyfication unit DU is coupled to the input of the fuzzyfication unit FU and/or to the input of the processing unit RU.

Abstract: Analog processor of antecedent parts of fuzzy logic inference rules and comprising a plurality of analog generators of membership function each having an output supplying a value corresponding to a degree of truth complemented to one (.alpha.') of logical assignments of the type (A is A') with the outputs being connected together to form a common circuit node and also connected to a current generator and the processor comprising also a voltage control device inserted between a supply voltage pole and a ground voltage reference and a one-way element connected to the common circuit node and the one-way element having an output producing an overall degree of truth for the antecedent part of the fuzzy rule to be processed.

Abstract: A circuit (14) and method which determine a degree of membership of an input in a membership set during a fuzzy logic operation. The degree of membership is calculated by a single "MEM" software instruction. The MEM instruction determines whether the system input has a degree of membership of zero, of a saturation level, or of some value in between. If a membership function has a boundary with an infinite slope, the slope value is set equal to zero and the degree of membership is then set equal to a saturation value for those input values corresponding to the infinite slope boundary.

Abstract: On the basis of a device without hidden neurons, arbitrary samples are taken from a set of learning samples so as to be presented as objects to be classified. Each time if the response is not correct, a hidden neuron (H.sub.i) is introduced with a connection to the output neuron (O.sub.j) of the class of the sample, whereas if the response is correct, no neuron whatsoever is added. During this introduction phase for hidden neurons, the neurons are subdivided into groups by searching for each of the introduced neurons whether it falls within an existing group, in which case it is incorporated therein; otherwise a new group is created around this neuron. The association with a group is defined as a function of the distance from the "creator" neuron. Application: character recognition systems.

Abstract: A new data structure describes an inheritance network with exceptions, augmented with rules attached to nodes in the network, a background context of rules and information, and an optional prioritization of links in the network. A process that determines which rules apply to classes (nodes) in the taxonomy exploits the structure of the network, in particular specificity and path prioritization, to get preferred maximally consistent subsets of rules that apply at specified nodes.

Abstract: A circuit for implementing a neural network comprises a one dimensional systolic array of processing elements controlled by a microprocessor. The one dimensional systolic array can implement weighted sum and radial based type networks including neurons with a variety of different activation functions. Pipelined processing and partitioning is used to optimize data flows in the systolic array. Accordingly, the inventive circuit can implement a variety of neural networks in a very efficient manner.

Abstract: Method and apparatus of parallel processing of multiple inference rules organized in fuzzy sets or logical functions of multiple fuzzy sets including membership functions defined in a so-called universe of discourse. The inference rules are configured essentially as IF-THEN rules with at least one antecedent preposition and at least one consequent implication. The prepositions have at least one term of comparison between membership functions and a plurality of input data and each term is separated by logical operators. The method associates with the logical operators maximum and minimum operations among two or more elements and calculates exhaustively the overall degree of truth (.OMEGA.) of a rule with a maximum or minimum of N partial truth levels. The method is accomplished by a plurality of identical, parallel inferential processors. Each inferential processor determines a preposition or a partial truth level of a preposition.

Abstract: Different bonus values and penalty values are allotted for each partial task of the unit such as, for example, drive from A to B, hold your positional uncertainty below a specific threshold, or draw up a map of the surroundings and add landmarks to it. Performance weightings for the individual tasks are yielded, in conjunction with a need to carry the latter out, after analysis of the bonus values and penalty values, and are evaluated in a control unit. Furthermore, in the context of the method a local planning horizon is specified in which the surroundings of the unit are subdivided into grid cells. Preferred directions, which lead the unit by the shortest path to already known or unconfirmed landmarks are stored for these grid cells, with the aim of reducing the positional uncertainty, or of being able to confirm a landmark. All the different routes which are possible within the framework of this grid are then investigated as to what contribution they make to enable the unit to reach the goal.

Abstract: An expert system method is provided that performs crime site analysis in a criminal activity area to determine a likely center of criminal activity. The method uses distances from related crime sites to determine a probability surface for the entire criminal activity area. A target area having highest values of probability becomes the center of criminal activity. The expert system can be used to access other crime databases such as witness reports to further define the criminals; residence or workplace.

Abstract: A cell for an electronic diffusion network that can be connected with neighboring cells for the transmission of information via linking elements (8n, 8rn, 8sn) having a nonlinear conduction characteristic. The cell also includes a connection node (1n) whose potential is representative of the excited state of the cell and to which is linked at least one of the linking elements (8n). The connection node (1n) is inserted into a series circuit which includes a current source (3n) and a current consuming element (5n) so as to fix the potential of the node (1n) to which the current source (3n) is connected by way of a switch element (2n) that has selective closure with the aid of activation pulses. A network equipped with such cells can assign both a temporal parameter and a quantitative parameter to the information diffused.

Abstract: In a parallel multi-value neural network having a main neural network 16 and a sub neural network 18 coupled with the main neural network 16 in parallel for an input signal, the main neural network 16 is trained with a training input signal by using a main multi-value teacher signal, and the sub neural network is successively trained with the training input signal by using multi-value errors between a multi-value output signal of the main neural network 16 derived through a multi-value threshold means 17 and the main multi-value teacher signal, so as to compensate the multi-value errors involved in the multi-value output signal of the main neural network 16 by the multi-value output signal of the sub neural network 18 derived through a multi-value threshold means 19. A desired multi-value output signal of the parallel multi-value neural network 15 is obtained by adding in modulo the multi-value output signals of both the neural networks through a multi-value modulo adder 20.

Abstract: A method and apparatus for the prediction of time series data, specifically, the prediction of a foreign currency exchange rate. The method disclosed transforms the time series data into a difference of a series, compresses the transformed data using a log transformation, converts the compressed data into symbols, and subsequently trains one or more neural networks on the symbols such that a prediction is generated. Alternative embodiments demonstrate the conversion by a self-organizing map and training by a recurrent neural network.