1,720,979 research outputs found
Intelligent Collision Avoidance, Control and Guidance
No full textThis paper reviews a variety of advanced signal processing algorithms that have been developed at the University of Southampton as part of the Prometheus (Programme for European traffic flow with highest efficiency and unprecedented safety) programme to achieve an intelligent driver warning system (IDWS). The IDWS includes the detection of road edges, lanes, obstacles and their tracking and identification, estimates of time to collision, and behavioural modelling of drivers for a variety of scenarios. The underlying algorithms are briefly discussed in support of the IDWS
An intelligent driver warning system for vehicle collision avoidance
No full textThis paper describes a basic architecture of an intelligent driver warning system which embodies an adaptive driver model for indirect collision avoidance. In this study, the driver modelling objective is focused only on longitudinal car-following, and the model inputs are chosen to be the past history of throttle angle, controlled vehicles's speed, range and range rate to the front vehicle whereas the model output is chosen to be the current throttle angle. An artificial neural network called CMAC and a conventional linear model are independently applied to model the real driver data taken from test track and motorway environments. The CMAC model is chosen because of its nonlinear modeling capability, on-line learning convergence and minimum learning interference characteristics, whereas the linear model is chosen as a control benchmark to examine the nonlinear characteristics of the driver's behaviour. The modeling capabilities are then evaluated based on one-step ahead prediction error performances over the training and testing sets, learning curves and correlation based model validation techniques. Modeling results suggest that the past history of the throttle dynamics plays a critical role in reducing the deviation of the error correction, which in turn suggest that the throttle dynamics is generally slow for road driving. Also, the time scale dependency of the model on the driver's behaviour varies significantly from the test track to motorway environment. In the driver modeling experiment, the time scale was chosen such that the deviation of the error correlation was minimised. The test track results suggest that the chosen inputs are indeed relevant variables for modeling the driver's behaviour. Unlike that of the CLM, the degree of error deviation was found to be acceptable for the test track scenario, implying a significant nonlinear coupling of the throttle output with the speed, range and range rate data. Whereas for the motorway data, the modeling performance for both models is comparable, and the time scale of the driver model is approximately three times longer than that used in the test track data
A Global Gradient Noise Covariance Expression for Stationary Real Gaussian Inputs
No full textSupervised parameter adaptation in many artificial neural networks is largely based on an instantaneous version of gradient descent called the Least-Mean-Square (LMS) algorithm. As the gradient is estimated using single samples of the input ensemble, its convergence properties generally deviate significantly from that of the true gradient descent because of the noise in the gradient estimate. It is thus important to study the gradient noise characteristics so that the convergence of the LMS algorithm can be analyzed from a new perspective. This paper considers only neural models which are linear with respect to their adaptable parameters, and has two major contributions. Firstly, it derives an expression for the gradient noise covariance under the assumption that the input samples are real, stationary, Gaussian distributed but can be partially correlated. This expression relates the gradient correlation and input correlation matrices to the gradient noise covariance, and explains why the gradient noise generally correlates maximally with the steepest principal axis and minimally with the one of the smallest curvature, regardless of the magnitude of the weight error. Secondly, a recursive expression for the weight error correlation matrix is derived in a straightforward manner using the gradient noise covariance, and comparisons are drawn with the complex LMS algorithm
A Stability Analysis of the Modified NLMS Rules
No full textThis paper investigates the stability of two recently proposed modified NLMS learning rules that are based on calculating the smallest weight change which stores the current training pattern exactly. The Lp (p = 1, 2, infinity) norm used to measure the weight update produces different learning algorithms, and it is shown that both new learning rules (p = 1, infinity) can become unstable, as the parameter error increases without bound. This is in direct contrast to the standard (p = 2 norm) NLMS rule which is unconditionally stable (in the sense described in this paper - monotonically non-increasing weight error), and indeed the NLMS rule was originally derived to overcome such limitations. The conditions under which instability can occur are investigated both theoretically and in simulation and are shown to depend on the form of the input vector and only indirectly on the learning rate
Design Improvements in Associative Memories for Cerebellar Model Articulation Controllers (CMAC)
No full textA number of recent improvements to the design of associative memories for CMAC systems are described. These are (i) an improved scheme for allocating C weights to a given input vector in Rn, (ii) design of receptive field shapes within the hypercube associated with an individual weight (including some experimental evaluations of these shapes), (iii) matching the field shapes to the hypercube itself using the concept of superspheres, (iv) speeding up the convergence of the weight training procedure
On Real Time Vehicle Guidance and Driver Modelling Within Prometheus
No full textThis paper presents our current driver modeling and vehicle guidance results for use within Prometheus. Both applications are implemented using B-Splines because of their fuzzy interpretability, fast learning convergence and minimum learning interference. The driver modeling results suggest that the past history of throttle angle tends to dominate, as indicated by its bias and lag in the driver model. Also, the training data exhibit measureable nonlinearity, and the degree is expected to increase for more complicated scenarios. Whereas for the parking problem the generalization performance is improved for a range of initial conditions via model adaptation, initialization and edition
On the Condition of Adaptive Neurofuzzy Models
No full textLearning within fuzzy and neurofuzzy systems is becomingly increasingly important as researchers try to infer qualitative, vague information from quantitative, numeric data. The fuzzy representation of an adaptive neurofuzzy system is important both for initialisation and validation purposes, where a designer needs to interpret the knowledge stored in a network. Therefore it is important to study the convergence and rate of convergence characteristics of the parameters in a neurofuzzy model and investigate how this depends on the system's structure. This paper considers how the condition of the input fuzzy sets determines the convergence and generalisation abilities of the network and describes several new results about instantaneous least mean square training rules
A Unified Framework of Associative Memory Neural Networks for Adaptive Dynamic Modelling, Control and Estimation
No full textAspects of Instantaneous On-Line Learning Rules
No full textIn neural and fuzzy learning systems, instantaneous learning rules have often been proposed for use within on-line adaptive modelling and control schemes. However many aspects of this work remain unexplained or only partially known such as: how do these learning rules deal with singular systems? what happens when the data are inconsistent? how is on-line parameter convergence related to that of standard gradient descent rules? and is momentum beneficial to the parameter estimation procedure? This paper investigates all of these topics, suggests modifications to the basic procedures where necessary and describes some of the reformulations which have been previously proposed
On the Convergence Rate Performance of Normalized Least-Mean-Square Adaptation
No full textThis paper compares the convergence rate performance of the Normalized Least-Mean-Square (or NLMS) algorithm to that of the standard Least-Mean-Square (LMS) algorithm, which is based on a well known interpretation of the NLMS algorithm as a form of LMS via input normalization. With this interpretation, the analysis is considerably simplified, and the difference in the rate of parameter convergence can be compared directly by evaluating the condition number of the normalized input correlation matrix and the unnormalized one. The main contribution of this paper has two parts. Firstly, it derives exact condition number expressions for the normalized input correlation matrix using an arbitrary odd-order filter length with two distinct unnormalized eigenvalues; whereas the corresponding even-order NLMS condition numbers are shown to be bounded between their odd-order counterparts. These expressions require that the input samples be statistically stationary and zero-mean Gaussian distributed, and provide an important insight into the relative convergence performance of the NLMS algorithm as a function of the filter length to that of the LMS. Secondly, this paper provides a conjecture which set bounds on the NLMS condition number for any arbitrary number of distinct unnormalized eigenvalues, and this conjecture has been found to be consistent with extensive computer simulations. Given that the same maximum and minimum unnormalized eigenvalues but with varying power levels associated with intermediate eigenvalues, this bound suggests that the NLMS convergence rate decreases with the number of unnormalized eigenvalues with excessive power levels
- …
