Abstract: A system for responding to telephone queries for transit system information is disclosed in which the information received by transit information operators is used to initiate a pathfinding query, which results in a search using a heuristic method with Island Sets being performed in order to determine, for example, an itinerary which the operator can relate to the caller which answers the caller's query. A stopping rule, based on clock time, iteration count, openings of goal nodes, adjustments of goal node labels and the state of the transaction queue is also disclosed.

Abstract: A learning apparatus for use in a neural network system which has a plurality of classes representing different meanings. The learning apparatus is provided for learning a number of different patterns, inputted by input vectors, and classified in different classes. The learning apparatus is constructed by a computer and it includes a section for producing a plurality of output vectors representing different classes in response to an input vector, a section for obtaining a first largest output vector of all the output vectors, a section for obtaining a second largest output vector of all the output vectors, and a section for setting predetermined weights to the first and second largest output vectors, respectively, such that the first largest output vector is made larger, and the second largest output vector is made smaller. Furthermore, a section for determining a ratio of the weighted first and second largest output vectors, respectively, is included.

Abstract: The robot is started after a groove line of a workpiece to be welded is taught to the robot as a target movement locus for the distal end of a welding torch, by using a fixed coordinate system. A control unit of the robot calculates a position (Pi) of an individual interpolation point in the fixed coordinate system in accordance with data specified in a program (S5), and successively calculates corrected positions Pi' (Xi+.SIGMA.xr', Yi+.SIGMA.yr', Zi+.SIGMA.zr') of respective interpolation points by using a cumulative value of correction amounts (S8). The correction amounts are obtained through transformation of an output of a visual sensor, representing a deviation of a current moving position of the torch from the groove line in a predetermined coordinate system defined by the current position and a current moving direction of the torch end, into a correction amount in the fixed coordinate system.

Abstract: A fuzzy computer which adopts hardware arrangement to simultaneously execute the upper bound cut, the maximum value operation, and the calculation of the center of gravity by directing attention to that the membership functions are overlapped double or triple at the best with each other when fuzzy operation is carried out, or which adopts software arrangement capable of executing the processing equivalent to the above hardware arrangement by employing a general microprocessor system, then, the number of comparison operation in the upper bound cut and maximum value operation is reduced, accordingly, executing the upper bound cut, the maximum value operation, and the calculation of the center of gravity at higher speed.

Abstract: A method of processing information in an artificial neural network including a plurality of artificial neurons and weighted links coupling the neurons. In the method, those of the artificial neurons whose output values change by a value greater than a threshold value are selected. The output values of the selected neurons are calculated, and the influence which the changes in the output values of the selected neurons impose on the input values of the other artificial neurons is computed. The threshold value is changed such that an appropriate number of neurons are selected. The information processing in the artificial neural network is stopped when the threshold value decreased below a predetermined small value and the values output by all artificial neurons change by a value equal to or less than the threshold value.

Abstract: A method for increasing the accuracy of an analog neural network which computers a sum-of-products between an input vector and a stored weight pattern is described. In one embodiment of the present invention, the method comprises initially training the network by programming the synapses with a certain weight pattern. The training may be carried out using any standard learning algorithm. Preferably, a back-propagation learning algorithm is employed.Next, the network is baked at an elevated temperature to effectuate a change in the weight pattern previously programmed during initial training. This change results from a charge redistribution which occurs within each of the synapses of the network. After baking, the network is then retrained to compensate for the change resulting from the charge redistribution. The baking and retraining steps may be successively repeated to increase the accuracy of the neural network to any desired level.

Abstract: An information processing element for processing information with a function of neural network includes a semiconductor integrated circuit element portion comprising a plurality of neuron circuit regions constituting a neuron function among the neural network function, a molecular film element having a light-electricity function, provided on the circuit element portion, and the combination between the plurality of neurons is realized by utilizing a photoconductivity property of the molecular film element.

Abstract: Disclosed is a control robot having a machining tool comprising a rotatable tool for grinding at the distal end of the robot arm thereof so as to carry out grinding work with pressing the machining tool against the surface of a work to be machined under predetermined pressure, comprising a posture control shaft for controlling the posture of the machining tool provided at the robot arm; a rotation shaft of the rotatable tool: and arrangement in which the posture control shaft and the rotation shaft of the rotatable tool are respectively arranged in different axial directions.

Abstract: A parallel processing computer system for clustering data points in continuous feature space by adaptively separating classes of patterns. The preferred embodiment for this massively parallel system includes preferably one computer processor per feature and requires a single a priori assumption of central tendency in the distributions defining the pattern classes. It advantageously exploits the presence of noise inherent in the data gathering to not only classify data points into clusters, but also measure the certainty of the classification for each data point, thereby identifying outliers and spurious data points. The system taught by the present invention is based upon the gaps between successive data values within single features. This single feature discrimination aspect is achieved by applying a minimax comparison involving gap lengths and locations of the largest and smallest gaps. Clustering may be performed in near-real-time on huge data spaces having unlimited numbers of features.

Abstract: A frequency-based neural network in which the state of a neuron is indicated by the frequency of an impulse stream emitted by the neuron uses an interconnectivity structure employing a frequency-modulation multiplexing scheme to weight and communicate the pulse stream from an emitting neuron to receiving neurons in another network level.

Abstract: A neural network architecture has phase-coherent alternating current neural input signals. Each input v.sub.k.sup.in is a two-phase pair of signals 180 degrees out of phase. Capacitive coupling of both signals of n input pairs to a summation line gives a non-dissipative realization of the weighted sum ##EQU1## with general real neural weights w.sub.ik. An alternating current offset signal proportional to u.sub.i is also capacitively coupled to the summation line. The signal on the summation line is passed through a low input capacitance follower/amplifier, a rectifier and a filter, producing a direct current signal proportional to the magnitude ##EQU2## This signal is compared with a direct current threshold proportional to t.sub.i, and the resultant is used to gate a two-phase alternating current output signal. The output is therefore functionally related to the inputs by ##EQU3## with .theta. the Heaviside step function.

Abstract: This invention is a novel high-speed neural network based processor for solving the "traveling salesman" and other global optimization problems. It comprises a novel hybrid architecture employing a binary synaptic array whose embodiment incorporates the fixed rules of the problem, such as the number of cities to be visited. The array is prompted by analog voltages representing variables such as distances. The processor incorporates two interconnected feedback networks, each of which solves part of the problem independently and simultaneously, yet which exchange information dynamically.

Abstract: A method and apparatus for the design of a robust vector quantizer is disclosed. The initial output vector set is equal to the centroid of a training sequence of input vectors. A neural-network simulation and neighborhood functions are utilized for splitting and optimizing the output vectors. In this manner, the entire output vector set is sensitive to each input vector and therefore optimal output vector locations with respect to specified distortion criteria are obtained. The resulting vector quantizer is robust for the class of signals represented by the training sequence.

Abstract: A vacuum cleaner comprising a fuzzy inferring device for determining a motor speed in response to the amount of dust, and means for holding the motor speed determined by the fuzzy inferring device for a predetermined period of time, whereby, after the motor speed is held for the predetermined period time, the motor is driven for a certain period of time at the speed subsequently determined by the fuzzy inferring device.

Abstract: Method and system of predicting attribute value in the learning system, (a) where expected value of equivalent example and its classification class is generated by domain knowledge, with respect to the test example being supplemented with candidates of a plural number of attribute values, in which the attribute lacking part of attribute values can adopt in the learning system, having (i) attribute descriptions containing a plural number of training examples, each of which contains attribute and its attribute value, and such training example and classification class adopted by each training example, and (ii) domain knowledge generating a pair of second example and its classification class from a pair of first example and its classification class, as equivalent example, (b) where classification class of equivalent example generated as above is predicted by decision tree produced based on the above attribute descriptions, and (c) where the attribute value lacking the above test example is predicted by use of the

Abstract: In an expert system of an operating system e.g. an industrial plant it is desired to input data in the form of relationships between operating parameters of the system to a database. However, if such data are input by inexperienced users, there is the possibility of incorrect data being input. Therefore, a memory stores a plurality of fundamental rules, each being a predetermined relationship between the operating parameters. Then when a datum is input, the fundamental rule(s) corresponding to the operating parameters of the datum are investigated, and the datum is transmitted to the database when each such corresponding fundamental rule is satisfied. In a further development, prevent transmitting datum to a database when each such corresponding fundamental rule is not satisfied.

Abstract: Dynamic control of the motion of a plurality of robots (12a and 12b) is accomplished by checking for interference if each robot were to traverse an incremental portion of its programmed path segment, the incremental portion being at least as long as the distance required for the robot to stop. Should there be no interference, each robot is allowed to traverse the incremental path length portion. Otherwise, the robot having the highest priority is enabled to reverse its incremental path segment portion while the other robots which would otherwise interfere are ordered to take evasive action, either to stop if moving, or to move away if stopped. Periodic checks are made for an interference condition at a sufficiently rapid rate so that each non-interfering robot can traverse successive incremental path segment portions in a smooth manner.

Abstract: A method for adaptively setting analog weights in analog cells of a neural network and the like. The process starts by addressing a synapse cell in the network. A target weight for said addressed synapse cell is selected, and the current weight present on the synapse cell is measured. The amplitude and duration of a voltage pulse to be applied to said synapse cell to adjust said synapse cell in the direction of said target weight is calculated using a set of coefficients representing the the physical characteristics of the synapse cell. The voltage pulse is applied to the addressed synapse cell and the new weight of the synapse cell is re-measured. If the synapse cell weight is within acceptable limits of the target weight, the values of the coefficients are saved and the next adjacent synapse cell is addressed until all synapse cells are set. If the synapse cell is not within acceptable limits, new values for the coefficients are calculated in relation to the re-measured weight.

Abstract: Digital synapse input signals which are temporally aligned with each other are converted to analog form for application to a neural net layer. Several digital synapse input signals can be written into digital memory at different times, to be read out in temporal alignment for conversion to analog form for application to a neural net layer. The digital synapse input signals can originate from apparatus for raster-scanning a field or successive fields of video samples, for example, with the neural net layer providing filters for recognizing shapes in a field of the video samples. The neural net layer can be either a pre-programmed type or a learning type, and its axonal outputs may be digitized for use in subsequent digital processing.

Abstract: A robot is controlled to move from a predetermined rest point to contact a moving target. The control parameters of the robot are not known. Control is effectuated by predetermining the time required for the robot to move from a rest position to each of a plurality of points within the region in which the target is expected to be found. When the target is identified, its location near the predetermined points at a future time is computed. The known time required from motion of the robot from the rest position to predetermined points adjacent the future location of the target is determined by reference to stored information. The exact time required for the motion may be interpolated between the predetermined points. The time for robot motion is subtracted from the future time to determine the time at which robot motion begins. In a particular embodiment of the invention, a robot coacts with a circular conveyor belt to retrieve objects from the conveyor and transport them to a dump point.