This paper presents a fast algorithmic method to train convolutional neural network (CNN) classifiers through extreme learning which has been verified on popular datasets on classification and pedestrian detection. CNN has been one of the best classifiers for images and object recognition. However, the Backpropagation (BP) algorithm, mostly used for training CNN, suffers from slow learning, local minimum, and poor generalization. To solve these problems, a novel architecture called CNN-ELM has been proposed here. Its core architecture is based on a local image (local receptive field) version of the ELM (Extreme Learning Machine) adopting random feature learning. Using MATLAB 2015a, classification experiments using the raw image data as input, show a comparable or mostly better classification performance compared to the BP trained CNN, with its training speed up to 200 times faster on MNIST, NORB, and CIFAR-10 datasets. Pedestrian detection experiments using INRIA datasets also exhibits much faster training than the BP trained CNN without sacrificing detection performance.

In this paper, we focused on designing a semi-rotation invariant feature descriptors for classification problem. We proposed hierarchical Zernike moments architecture which is combination of original Zernike moments, the local receptive field concept and shared weights concept from convolution neural network. The descriptors which are output of the architecture have improved classification performance while maintaining rotation invariant property within semi-rotation region. Experimental results for classifying designed descriptors, extracted from semi-rotated MNIST dataset, with multi-layer perceptron demonstrate the effectiveness of the proposed method compared with other methods, such as original Zernike moments and convolution neural network.

We introduce a facial expression training system using the bilinear shape model which helps people to practice making a facial expression. The user face on the camera preview screen is reconstructed into a 3D face model and the model is transformed to blend shape model which represents the facial expression. This way, the system can precisely analyze the facial expression of the user. With target 3D face model appearing on the screen, the 3D face model changes its facial expression, it leads the user to change his facial expression to become look like same. The system recognizes whether the facial expression of the user is same with the one of 3D face model. As the system gives the various missions to user to change his facial expression, user can practice the facial expression. It can be used for bell's palsy patient who needs face rehabilitation exercise or someone who need to practice unique facial expression such as stewardess smile or facial mimicry.

We present ARgo(Augmented Reality tamaGOchi), a system for animating an everyday object on a smart device. We focus to interact with every everyday object on the smart device augmented reality(AR) through animating them with animating and gaming contents. We developed a Single-Point-Tap gesture event based object tracking algorithm using a GrabCut, TLD and CamShift tracking hybrid system for a markerless mobile AR application. To make the object seem to have an inherent facial expression, we use Poisson image edit, which is a seamless blending approach. To instill 'life' to an everyday object, we adopted the Tamagotchi storytelling concept. For multi-platform game application purposes and computer vision programming, the AR programming environment was built using the Cocos2d-x game engine and OpenCV. We show the use of our system to animate the everyday object in real environment.

In this paper, we propose a new ceiling vision-based active mapping and exploration framework for wide-open structural environments such as museum, exhibition center. We adopt the ceiling vision based approach with model-free landmark definition and visual servoing based data association. Additionally, by using a recursive explore and exploit strategy in topological map-building and navigation, we could gradually spread out the mapped region in unknown, wide-open environments without degrading map accuracy. Simulations and real experiment demonstrated that our mapping and exploration algorithm could work well in real environment.

This paper presents an efficient approach to solve the problem of real-time robust hand segmentation and fingertip extraction for finger gesture recognition in Human-robot interaction (HRI). Firstly, we propose an improved cascade filtering based hand region candidate estimation by the combination of YCbCrskin color probability image based segmentation with genetic algorithm and post-processing with morphological operation and blob analysis in blurred, low-resolution images. Using the segmented hand candidate regions, we estimate the hand region center and fingertip position from distance transform and geometrical feature of hand. From the hand orientation and hand/palm center, we find the optimal each fingertip position and its orientation. Experimental results show that the proposed algorithm not only rapidly detects the hand regions under various illumination conditions, but also it efficiently extracts the finger information with size and rotation invariance.

Three-dimensional (3D) data processing has recently acquired greater importance in solving complex tasks such as object recognition, environment modeling, and robotic mapping and localization. Since using raw 3D data without preprocessing is very time-consuming, extraction of geometric features that describe the environment concisely is essential. A plane is a suitable geometric feature due to its richness and simplicity of extraction. This paper presents an online incremental plane extraction method using line segments. Since our system is based on a nodding laser scanner, we exploit the incremental nature of data acquisition in which physical rotation and algorithm implementation are conducted in parallel. In contrast to other plane extraction methods, line segments defined by two end points become supporting elements that comprise a plane, so we need not handle all the scan points once the line segments are extracted from each scan slice. This reduces the algorithms complexity and the computation time. Experimental validation and comparison with state of the art method were conducted using tens of complete scan data sets acquired from a typical indoor environment.

We introduce a fast Nearest Neighbor Search (NNS) algorithm using an Approximate Cached k-d tree (ACk-d tree) structure for low dimensional data sets. The search process of the standard k-d tree starts from the root node and employs a tentative back-tracking search. In contrast, the proposed method begins to search at the appropriate leaf node (cached node) and applies a depth-first nontentative search. This method improves searching speed, with tradeoff of the searching accuracy. To get a proper starting node, the proposed method is based on two properties: i) The ithquery point is likely to be close to the (i-1)thquery point, ii) The ithquery point is likely to be close to the ithmodel point. These properties are rather right, in case of practical 3D point sets which are consecutively acquired from 3D point sensors (e.g. a stereo camera, the Kinect sensor, and LIDAR). Results show that the search time of the proposed method is superior to other variants of k-d tree for practical point data sets.

This paper presents a fast and robust plane feature extraction and matching technique for RGB-D type sensors. We propose three algorithm components required to utilize the plane features in an online Simultaneous Localization and Mapping (SLAM) problem: fast plane extraction, frame-to-frame constraint estimation, and plane merging. For the fast plane extraction, we estimate local surface normals and curvatures by a simple spherical model and then segment points using a modified flood fill algorithm. In plane parameter estimation, we suggest a new uncertainty estimation method which is robust against the measurement bias, and also introduce a fast boundary modeling method. We associate the plane features based on both the parameters and the spatial coverage, and estimate the stable constraints by the cost function with a regulation term. Also, our plane merging technique provides a way of generating local maps that are useful for estimating loop closure constraints. We have performed real-world experiments at our lab environment. The results demonstrate the efficiency and robustness of the proposed algorithm.

The Iterative Closest Point (ICP) algorithm is one of the most popular methods for geometric alignment of 3-dimensional data points. We focus on how to make it faster for 3D range scanner in intelligent vehicle. The ICP algorithm mainly consists of two parts: nearest neighbor search and estimation of transformation between two data sets. The former is the most time consuming process. Many variants of the k-d trees have been introduced to accelerate the search. This paper presents a remarkably efficient search procedure, exploiting two concepts of approximate nearest neighbor and local search. Consequently, the proposed algorithm is about 24 times faster than the standard k-d tree.

This paper presents a novel solution to overcome the disturbance noise (outlier) for the Extended Kalman Filter based Simultaneous Localization And Mapping (EKF-SLAM). The standard Kalman Filter (KF) is not robust to the disturbance noise. The possibility that disturbance may happen is high, because SLAM aims at exploring unknown environment. Hence KF based SLAM methods should consider how to handle the disturbance noise. Variations of KF have been introduced to overcome this problem. However, these methods employ manual parameter tuning, detecting/weighting method. The core of our algorithm is to inflate the state uncertainty by using the magnitude of innovation, without tuning and detecting. Although it is impossible to estimate the state value immediately, the inflated state uncertainty makes it possible for the estimated value to converge on the true value much faster. We evaluate the proposed method under the well-known benchmark Matlab program. The results show that the proposed method overcomes the disturbance noise and increases the performance of EKF-SLAM.

In this paper, we develop a new framework for simultaneous localization and mapping (SLAM) based on Rao-Blackwellized particle filters (RBPF), that can be applied to floor-cleaning robots which are equipped with sparse and short-range sensors. To overcome the sensor limitations, the entire region is divided into several local maps, which are assumed to be independent to each other. The local maps are estimated by a local RBPF SLAM, and then the trajectory of the local map origin is estimated by a global RBPF SLAM. To compensate for the severe sensing limitations, we also adopt the assumption that the indoor environments consist of many orthogonal lines. This assumption significantly enhances the filter performance. The proposed SLAM framework is combined with a coverage path planning algorithm, and the resulting robot system is capable of online simultaneous coverage and SLAM. The algorithm was embedded into a real mobile robot platform and tested in a real home environment to assess the robustness of the proposed method.

This article presents an efficient visual landmark extraction and recognition method that can autonomously and rapidly detect visual features such as objects or groups of small objects, and that can be applied to visual object recognition based SLAM and navigation in indoor/large environments using a monocular/omnidirection vision system. Our method consists of two-stage: (1) we autonomously extract object regions with modified fuzzy object segmentation. We generate a saliency map of the scene based on Modified Phase spectrum of Fourier Transform (mPFT) and extract the final salient object landmark with weighted combination of candidate of objects and saliency map. (2) Using these result, we register current objects as visual landmark and then recognize the current image the scale invariant feature transform (SIFT) - based recognition with probabilistic voting. In experiments results in real indoor and large hall environments, the proposed method was simpler and 10~15% better performance in computation efficiency and successfully extracted salient object landmark in complex environments with high recognition rates. The proposed algorithm can be easily implemented in real-time by reducing the number of objects considered.

This paper proposes a method for a cleaning robot to generate a topological map from an occupancy grid-map. Virtual door is defined as the candidates of real door, and the virtual doors are detected as edges of the topological map by extracting corner features from the occupancy grid-map; using this method, an initial topological map is generated, which consists of nodes and edges. The final topological map is generated using a genetic algorithm to merge the nodes and reduce the edges. As a result, the generated topological map consists of nodes divided by virtual doors and edges located in real doors. The proposed methods provide a topological map for the user interaction and the cleaning robot, and the topological map can be used to plan more efficient motion including room-to-room path planning and covering each room.

This paper proposes a method for reconstructing non-rigid 3D shapes from noisy 2D shapes. The proposed method estimates the 3D shape bases and projection matrices, exploiting low-rank constraints. Then the method finds the optimal coefficients for linear combinations of 3D shape bases to represent non-rigid 3D shapes using a genetic algorithm, and refines the 3D shape bases and the projection matrices using gradient descent techniques. The method reconstructed correct non-rigid 3D shapes in the presence of noise. The results can be used in many areas including animation, motion capture and non-rigid 3D object tracking.

This paper presents a simultaneous localization and mapping (SLAM) of a large indoor environment using Rao-Blackwellized particle filter (RBPF) along with line segments as the landmarks. To represent the environment as a compact form, we use only two end points of the line segment, reducing computational cost in modeling line uncertainty. With a modified scan point clustering method, the proposed adaptive iterative end point fitting (IEPF) plays an important role in estimating line parameters by taking a noisy scan point near end points into account. Thus, by line-segment matching the robot is localized well in a local frame. We also introduce an online global optimization of a map, which provides more consistent map by removing spurious lines and merging collinear lines. Each of our approaches is efficiently integrated into the proposed RBPF-SLAM framework. Experiments with well-known data set demonstrate that the proposed method provides a reliable SLAM performance along with a compact map representation.

This paper presents an on-line complete-coverage path planning algorithm for mobile robots based on approximate cellular decomposition, which abstracts the target environment using grid. Most existing grid-based coverage algorithms have a common problem of constrained mobility which degrades the efficiency of the coverage task by inducing zigzag like path. In this paper, we propose a new complete coverage path generation algorithm, linked-smooth-spiral-path (LSSP), which removes the constraint on mobility by adopting a high-resolution grid-map representation of the environment and a cardinal-spline curve-model to generate a smooth spiral coverage path. We define a new performance measure for coverage tasks, which quantifies the smoothness of the coverage path. Simulation results demonstrate the improved coverage performance of the proposed algorithm compared to other existing grid-based coverage algorithms.

Non-linearity of color changing in various lighting conditions is one of the primary factors which make lane color recognition difficult. This paper introduces an illumination invariant lane color recognition method which can recognize two lane colors (white, yellow) and copes with the non-linearity by using neural networks. Our method utilizes the road texture as the indicator of illumination condition and learns the relation between illumination condition and lane color by using multi-layer perceptron. The proposed method was verified by the experiment with a road images sequence of real driving situation.

In this paper, we address a method for improving accuracy of a Neural Network (NN) aided Extended Kalman Filter (EKF) based SLAM by compensating for an odometric error of a robot. The NN is used for estimating the odometric error and online learning of NN is implemented by augmenting the synaptic weights of the NN as the elements of state vector in the EKF-SLAM process. Due to this trainability, the NN could adapt to systematic error of the robot without any prior knowledge and the proposed NN aided EKF-SLAM is very effective compared to the standard EKF-SLAM method under the colored noise or systematic bias error. Experimental results are presented to validate that our NN aided EKF-SLAM generates more accurate feature map than conventional EKF-SLAM.

In this paper, we address a method for improving the accuracy of the feature map from the extended Kalman filter based SLAM (EKF SLAM) by estimating the systematic parameters of the robot. Most error of the robot while traveling is divided into two categories: systematic and non systematic error. The systematic error contributes much more to odometry errors than non systematic one on most smooth indoor surfaces. So, we appended the systematic parameters of the robot to the state vector of EKF SLAM as its elements and estimated the systematic parameters while performing the prediction and update state of EKF SLAM. Because the additional elements to be estimated are appended to the state vector of the EKF SLAM, this is called an augmented EKF SLAM (AEKF SLAM). Experimental result is presented to validate that our AEKF SLAM is able to generate a more accurate feature map than conventional EKF SLAM by decreasing odometric error of the robot.

This paper proposes a method for detecting vehicles in urban traffic. The proposed method extracts vehicle candidates using AdaBoost. The candidate extraction process was speeded up further, exploiting inverse perspective transform matrix. Then the vehicle candidates were verified by the existence of vertical and horizontal edges. The detected vehicle regions were corrected by the vertical edges and shadow. Our algorithm showed the detection rate of 90.77% in urban traffic under normal lighting condition. The proposed algorithm can also detect vehicles in heavy rain. Our algorithm takes 37.13ms on average to detect vehicles in 320 by 240 images on a laptop computer (Intel ® Core™2 T7200, 2.00GHz, 1.00GB RAM).

This paper addresses a solution of simultaneous localization and mapping (SLAM) using a neuro evolutionary optimization algorithm. The proposed algorithm solves the global optimization problem of the SLAM using the cost function which represents the quality of a robot's trajectory in a world coordinate frame. In our algorithm, the neural network helps to estimate the difference of two consecutive positions accurately using the sensor inputs at each position. The evolutionary programming (EP) is used to find the most suitable neural network. The proposed neural network based SLAM is applied to the robot which has sonar sensors. The various experimental results demonstrate that the neural network based SLAM guarantees a consistent environmental map under the sonar sensor measurements.

This paper addresses the measurement noise of Extended Kalman Filter-based Simultaneous Localization And Mapping (EKF-SLAM). The Extended Kalman Filter (EKF) is based on the Gaussian noise with zero mean and should know the correct prior knowledge of control and measurement noise covariance matrices. If these conditions are not satisfied, EKF unavoidably diverges. The present paper proposes the method of a new adaptive kalman filter to be supported by Measurement Noise Estimator (MNE), which estimates the measurement noise distribution including biased noise and noise covariance, whenever the update step executes. We evaluate this method under well-known benchmark environment for SLAM problem. Simulation results show that the proposed algorithm overcomes degrading performance of the standard EKF under the condition of wrong knowledge of sensor statistics.

This paper presents a sensor-based online coverage path planning algorithm guaranteeing a complete coverage of unstructured planar environments by a mobile robot. The proposed complete coverage algorithm abstracts the environment as a union of robot-sized cells and then uses a spiral filling rule. It can be largely classified as an approximate cellular decomposition approach as defined by Choset. In this paper, we first propose a special map coordinate assignment scheme based on active wall-finding using the history of sensor readings, which can drastically reduce the number of turns on the generated coverage path. Next, we develop an efficient path planner to link the simple spiral paths using the constrained inverse distance transform that we introduced the first time. This planner selects the next target cell which is at the minimal path length away from the current cell among the remaining non-contiguous uncovered cells while at the same time, finding the path to this target to save both the memory and time which are important concern in embedded robotics. Experiments on both simulated and real cleaning robots demonstrate the practical efficiency and robustness of the proposed algorithm.

This paper introduces a pose invariant face recognition method with a training image and a query image using 3D morphable model and neural network. Our system uses 3D morphable model to get the reconstructed 3D face from the training image and obtains 2D image patches of facial components from the 3D face under varying head pose. The 2D image patches are used to train a neural network for pose invariant face recognition. Because those patches are obtained from the varying head pose, the neural network has robustness in the query image under the different head pose form the training image. Our pose invariant face recognition system has the performance of correct recognition higher than 98% with BJUT 3D scan database.

This paper develops a novel automatic parking algorithm based on a fuzzy logic controller with the vehicle pose for the input and the steering rate for the output. It localizes the vehicle by using only external sensors -a vision sensor and ultrasonic sensors. Then it automatically learns an optimal fuzzy if-then rule set from training data. This is possible using a genetic fuzzy system which optimizes the parameters for the fuzzy logic controller. Furthermore, it also finds the green zone of the ready-to-reverse position where parking is possible just by reversing. It has been tested on a Pioneer mobile robot which emulates the real vehicle.

This paper proposes an efficient human-like memory and memory management which utilizes Walsh-based distributed associative memory in reducing the computer storage and processing for pattern recognition. As a verification example, a memory storing 26 binary alphabet images takes only the physical space needed to store 8 patterns and yet capable of perfect recognition. Further, the experimental results show that the proposed memory management strategy can nicely deal with data transfer from short-term (working) memory to long-term memory.

A Walsh function based distributed associative memory is capable of storing multiple patterns in a single storage space with Walsh encoding of each pattern. This Walsh-based associative memory has unique advantages in aspects of both reduced storage and fast recall compared to other types of associative memories. However, the price to pay for these incredible benefits is the amount of crosstalk among stored patterns that sometimes leads to mis-recognition of some of the stored items. In this paper, this adverse effect arising from the superimposed storage can be greatly alleviated by optimizing the different sequencies of the Walsh functions associated with each pattern to be stored. This optimization also lends itself to maximize the memory capacity by jamming more patterns onto the same memory space still maintaining perfect recognition. In order to verify its efficiency, we successfully applied the Walsh-based memory to high speed face recognition with much reduced data storage.

This paper presents a novel real-time SLAM method working in an unstructured indoor environment with a single forward viewing camera. Most existing visual SLAM methods extract features from the environment, associate them in different images and produce a feature map as a result. However, our approach estimates the distances between the robot and the obstacles by applying a visual sonar technique to the image and then associates this range data through an iterative closest point (ICP) based algorithm and finally produces the grid map. Here, we reconstruct a pseudo-dense scan pattern (PDSP) which is a temporal accumulation of the visual sonar readings based on odometry readings to overcome the sparseness of the visual sonar and then apply ICP based algorithm to associate this scan pattern with the previous one. In addition, we further correct the slight trajectory error included in the PDSP reconstruction step to get a much more refined map. Experimental results show that our method can obtain a grid map using a single camera alone unlike active sensor-based SLAMs

Linear associative memory (LAM) has two serious problems to be practical. One is a large space of memory. The other is a high computational complexity. In this paper, we propose partitioning of the input data to alleviate these problems. The optimal linear associative memory (OLAM) minimizes the error of recall employing a pseudo-inverse operation. The proposed algorithm was applied to auto-associative recall of facial images. We show that the proposed algorithm can both reduce the memory space and computational complexity over the conventional optimal linear associative memory.

This paper proposes a technique for real-time recognition of facial expression which uses the active appearance model (AAM) with a second order minimization and a neural network. The second order minimization gives the ability of correct convergence with a little loss of frame rate to AAM. And the correctly extracted facial shape with AAM prevents the recognition of facial expression from undergoing a large error. In addition, high dimensional feature vectors of six facial expressions which consist of facial shape and texture can be dealt by the multi-layer perceptron, a type of the neural network, with a very high recognition rate of over 99%.

We present a system that calculates the position and the orientation of the robot using camera in the real time. There are several methods that calculate parameters of the relation between the world coordinate and the camera coordinate in the image processing area. Instead of the method using the parameteric way, our method obtains the robot pose by using the neural network. In this process, we need to remove the distortion of the pinhole camera image resulting in pose error orientation in the camera image. The neural network essentially spread the distorted image. Our method exhibits high accuracy with simple calculations.

This paper proposes a fast face detection technique which can find exact face regions in both gray and color static images using the Bayesian discriminating feature and the particle attractive genetic algorithm. In Bayesian discriminating feature method, face and nonface probability can be calculated with probabilistic models of the face and nonface feature vectors which consists of horizontal, vertical histograms, and 1D wavelets of the image inside the candidate window. These probabilities are modeled as Gaussian distribution and can be used to distinguish face regions from nonface regions. To search proper face candidate regions, we propose the particle attractive genetic algorithm which can fast converge on the exact face region. The proposed method demonstrates fast and precise detection results with the images obtained from offices or outdoor environments.

Active research into vehicle detection and tracking using a vision sensor are done for driver assistance systems (DAS) - collision warning and avoidance, vision enhancement, etc. The vehicle detection and tracking algorithm for DAS requires a robust feature extraction and tracking method regardless of the light and road conditions and an exact estimation of vehicle position and velocity regardless of the distance from the ego-vehicle. But most research was carried out in the day time with good lighting conditions and the little research done so far in the night time assumed no interference of headlights from other vehicles. This paper proposes a new robust vehicle detection and tracking method regardless of the light and road conditions at any distance using vision and sonar sensors. We use the sonar sensor for detection and distance estimation within 10 m and the image sensor over 10 m. First, this paper proposes a simple method that can determine the light condition by observing several images and this light condition is used by selecting one of several detection methods. The proposed vehicle detection method in the day time image can extract the shadow region represented by the boundary between a vehicle and the road and further verify using other vehicle features, such as symmetry rate, vertical edge, and lane information. The vehicle tracking method in the day time uses online template matching using the mean image created by several consecutive detection results. The vehicle detection method in the night time extracts bright regions caused by the headlights, taillights, brake lights, etc. and these candidates are verified by observing several consecutive frames.

Three-feature based automatic lane detection algorithm (TFALDA) is a new lane detection algorithm which is simple, robust, and efficient, thus suitable for real-time processing in cluttered road environments without a priori knowledge on them. Three features of a lane boundary - starting position, direction (or orientation), and its gray-level intensity features comprising a lane vector are obtained via simple image processing. Out of the many possible lane boundary candidates, the best one is then chosen as the one at a minimum distance from the previous lane vector according to a weighted distance metric in which each feature is assigned a different weight. An evolutionary algorithm then finds the optimal weights for combination of the three features that minimize the rate of detection error. The proposed algorithm was successfully applied to a series of actual road following experiments using the PRV (POSTECH research vehicle) II both on campus roads and nearby highways.

Evolutionary programming is mainly characterized by its mutation and selection rules. This paper first proposes a new mutation method that utilizes all the three generations in a family. A child regards its parent as a good or a bad teacher depending on whether it did better or worse than its grandparent. According to this outcome, the child creates a grandchild in the same or the opposite direction to the previous search direction. This 3-generation method allows the use of the gradient information in the search process. Second, contrary to the usual global competition, local competition which takes place in the subgroups of the whole population allows some inferior solutions with a good future potential to survive for enhanced diversity. Benchmark tests reveal the excellent performance of the proposed algorithm.

A local navigation algorithm for mobile robots is proposed that combines rule-based and neural network approaches. First, the extended virtual force field (EVFF), an extension of the conventional virtual force field (VFF), implements a rule base under the potential field concept. Second, the neural network performs fusion of the three primitive behaviors generated by EVFF. Finally, evolutionary programming is used to optimize the weights of the neural network with an arbitrary form of objective function. Furthermore, a multinetwork version of the fusion neural network has been proposed that lends itself to not only an efficient architecture but also a greatly enhanced generalization capability. Herein, the global path environment has been classified into a number of basic local path environments to which each module has been optimized with higher resolution and better generalization. These techniques have been verified through computer simulation under a collection of complex and varying environments.

A neural network based navigation algorithm is proposed for mobile robots using ultrasonic sensors. The neural network has a dynamically reconfigurable structure which can not only optimize the weights but also the input sensory connectivity in order to meet any user-defined objective. Further, in order to enhance generalization capability of a single network, a modular network is used in which each network module is optimized for a specific local environment based on environment classification. Both computer simulation and real experiments show the effective performance of the algorithm.

A local navigation algorithm for mobile robots is proposed, based on the new extended virtual force field (EVFF) concept, neural network-based fusion for the three primitive behaviors generated by the EVFF, and the evolutionary programming-based optimization of the neural network weights. Furthermore, a multi-network version of the above neurally-combined EVFF has been proposed that lends itself not only to an efficient architecture but also to a greatly enhanced generalization capability. These techniques have been verified through both simulation and real experiments under a collection of complex environments.

Evolutionary programming is mainly characterized by two factors: the selection strategy and the mutation rule. This letter presents a new mutation rule that has the same form as the well-known backpropagation learning rule for neural networks. The proposed mutation rule assigns the best individual's fitness as the temporary target at each generation. The temporal error, the distance between the target and an individual at hand, is used to improve the exploration of the search space by guiding the direction of evolution. The momentum, i.e., the accumulated evolution information for the individual, speeds up convergence. The efficiency and robustness of the proposed algorithm are assessed on several benchmark test functions.

In this paper, we propose a rule-level co-evolutionary approach using multiple subpopulations to evolve fuzzy logic controllers (FLCs). Each rule is used as the individual and the subpopulations each comprising a number of candidate rules co-evolve such that the rules belonging to the same subpopulation compete while those in different subpopulations cooperate to achieve the goal of finding a better FLC. During this process, the rules within each subpopulation become specialized into a kind of expert in the corresponding problem domain. For this approach, a simple credit assignment scheme for rule evaluation is introduced to effectively reduce the search space. The superiority of the proposed algorithm over traditional FLC-level evolution approach has been demonstrated by evolving FLCs for a typical nonlinear control problem-the ball and beam system.

The feasibility of using neural networks for camera localization and mobile robot control is investigated here. This approach has the advantages of eliminating the laborious and error-prone process of imaging system modeling and calibration procedures. Basically, two different approaches of using neural networks are introduced of which one is a hybrid approach combining neural networks and the pinhole-based analytic solution while the other is purely neural network based. These techniques have been tested and compared through both simulation and real-time experiments and are shown to yield more precise localization than analytic approaches. Furthermore, this neural localization method is also shown to be directly applicable to the navigation control of an experimental mobile robot along the hallway purely guided by a dark wall strip. It also facilitates multi-sensor fusion through the use of multiple sensors of different types for control due to the network's capability of learning without models.

A stable neural control scheme using a locally activated neural network has been proposed for a class of nonlinear dynamic systems. The locally activated neural network, for a given input, essentially selects a small subset of the network hidden nodes for output computation using the CMAC-like content addressing mechanism. This network aims to maintain local representations of the system dynamics. Thus, the global control performance in the concerned state space is achieved by the cooperation of many local control efforts and furthermore, real-time control can be facilitated because only a small sized network is involved to control and learn at any given time. The proposed control scheme is composed of two stages: (1) prediction error based learning in which the network attempts to learn the nonlinear basis functions of the plant inverse dynamics by a modified backpropagation learning rule; and (2) tracking error based learning in which the network weights are further fine-tuned using the basis set obtained in (1). This basis set spans the locally partitioned vector space of the system inverse dynamics when the prediction error based learning is achieved within a prescribed error tolerance. For uniform stability, the sliding mode control is introduced as a safety mechanism when the network has not sufficiently learned the plant dynamics yet. With suitable assumptions on the controlled plant, global stability and tracking error convergence proof has been given. Finally, the proposed control scheme is verified with computer simulation.

The industrial robot's dynamic performance is frequently measured by positioning accuracy at high speeds and a good dynamic controller is essential that can accurately compute robot dynamics at a servo rate high enough to ensure system stability. A real-time dynamic controller for an industrial robot is developed here using neural networks. First, an efficient time-selectable hidden layer architecture has been developed based on system dynamics localized in time, which lends itself to real-time learning and control along with enhanced mapping accuracy. Second, the neural network architecture has also been specially tuned to accommodate servo dynamics. This not only facilitates the system design through reduced sensing requirements for the controller but also enhances the control performance over the control architecture neglecting servo dynamics. Experimental results demonstrate the controller's excellent learning and control performances compared with a conventional controller and thus has good potential for practical use in industrial robots.

An indirect learning and direct adaptive control scheme based on neural networks-composite adaptive neural control-is proposed for a class of nonlinear systems. With indirect learning, the neural network learns the nonlinear basis functions of the system inverse dynamics by a modified backpropagation learning rule. The basis set spans the locally partitioned vector space of inverse dynamics with direct adaptation when indirect learning is achieved within a prescribed error tolerance. For localization of the state space of inverse dynamics, the hash addressing technique from CMAC is used for selecting only a small subset of the network hidden nodes according to where the input vector lies. As such, the global control performance can be obtained by the cooperation of many local convergence properties. For uniform stability, sliding mode control is introduced when the neural network has not sufficiently learned the plant dynamics. With suitable assumptions on the controlled plant, global stability and tracking error convergence proof can be given. Finally, the proposed control scheme is verified with computer simulation.>

The direct problem of dynamics for robotic systems with multiple simple closed chains has been formulated and solved. These systems include multiple robot manipulators performing assembly or cooperative tasks as well as multilegged walking machines. The basic problem in direct dynamics is to determine joint accelerations given the applied torques or forces and the system state. As such its solution will allow control designers to test and evaluate the performances of various control algorithms through computer simulation without having to build several costly prototypes. A unified approach is taken so that a single set of equations may describe dynamics of both multiple manipulators and legged vehicles.

As the cost of computer memory continues to decline faster than that of processors, it may be realistic to effectively apply pattern recognition methodology to security evaluation of an electric power system. Efficient implementation techniques are developed to achieve assessment in real time with a modest level of memory requirement. The basic idea is to recognize the unknown security of a particular system state operation from stored knowledge about similar operating patterns. Two efficient data structures are proposed here for its implementation. First, a distributed memory device, an associative memory is developed for recognition. This particular memory is found to be capable of parallel pattern matching along with reduced computer storage. Second, for an efficient implementation of the memory structure, these associative memories are configured in a hierarchical structure which not only expands storage capacity but also utilizes the speed of tree search. This structure provides a basis of an error-free, rapid, and memory-saving recognition algorithm.

A computer model for a distributed associative memory has been developed based on Walsh-Hadamard functions. In this memory device, the information storage is distributed over the entire memory medium and thereby lends itself to parallel comparison of the input with stored data. These inherent economic storage and parallel processing capabilities may be found effective especially in real-time processing of large amount of information. However, overlaying different pieces of data in the same memory medium creates the problem of interference or crosstalk between stored data and may lead to recognition errors. In this paper, a crosstalk reduction technique utilizing the gradient descent procedure is developed first. This minimizes the memory processing error and enhances memory saving. Second, for an efficient implementation of the memory structure, these associative memories are configured in a hierarchical structure which not only expands storage capacity but also utilizes the speed of tree search. Finally, a self-correcting technique is developed which achieves error-free recognition of near neighbors for any training pattern even among the presence of crosstalk.