Timothy Mark Feldkamp

Simple and conditioned adaptive behavior from Kalman filter trained recurrent networks

We illustrate the ability of a fixed-weight neural network, trained with Kalman filter methods, to perform tasks that are usually entrusted to an explicitly adaptive system. Following a simple example, we demonstrate that such a network can be trained to exhibit input-output behavior that depends on which of two conditioning tasks was performed a substantial number of time steps in the past. This behavior can also be made to survive an intervening interference task.

Neural network training with the nprKF

We present the training of feedforward and recurrent networks based on a more accurate treatment of the Kalman filter as applied to nonlinear systems.

A new approach to cluster-weighted modeling

We discuss an approach to joint density estimation called cluster-weighted modeling (CWM). The base approach was originally proposed by Gershenfeld (1998). We describe two innovations to the base CWM. Among these, the first enables the CWM to work with continuous streams of data. The second addresses the commonplace problem of local minima which may be encountered during the CWM parameter adjustment process. Our approach to mitigate this problem is quite elaborate, but it represents a principled way of improving the efficacy of the parameter adjustment process. We illustrate CWM and our performance enhancements with an example.

Custom VLSI ASIC for automotive applications with recurrent networks

Demands on the performance of vehicle control and diagnostic systems are steadily increasing as a consequence of stiff global competition and government mandates. Neural networks provide a means of creating control and diagnostic strategies that will help in meeting these demands efficiently and robustly. This paper describes a VLSI design that permits such networks to be executed in real time as well as the application in misfire detection, that served as a focus for the collaborative effort.

Signal Processing by Neural Networks to Create ``Virtual'' Sensors and Model-Based Diagnostics

This paper discusses the application of advanced neural network methods to the development of diagnostics for complex, nonlinear dynamical systems, for which accurate, first-principles models either do no exist or are difficult to derive. We consider two approaches to detect and identify failures in these systems. First, neural networks are trained to act as virtual sensors that emulate the performance of laboratory-quality sensors; this approach provides higher quality diagnostic information than is available directly from production sensors. Second, neural networks are trained to emulate nominal (fault-free) system behavior; model-based fault diagnosis is subsequently achieved by detecting significant deviations between actual and predicted system performance. We present experimental evidence of the viability of both approaches for a difficult automotive diagnostic task.

Conditioned adaptive behavior from Kalman filter trained recurrent networks

We demonstrate that a fixed-weight neural network can be trained with Kalman filter methods to exhibit input-output behavior that depends on which of two conditioning tasks had been performed a substantial number of time steps in the past. This behavior can also be made to survive an intervening interference task.

Ensembles of neural networks with generalization capabilities for vehicle fault diagnostics

This paper presents a two-step ensemble approach for vehicle fault diagnostics, an ensemble selection algorithm, BFES, and an analog Bayesian ensemble decision function, A-Bayesian-Entropy. We show through experiments that a neural network ensemble designed and trained by the proposed methodology, and selected by BFES with A-Bayesian-Entropy as the ensemble decision function can generalize well to vehicle models that are different from the vehicles used to generate training data.

Cluster-weighted modeling with multiclusters

Cluster-weighted modeling (CWM) was proposed by Gershenfeld (1999) for density estimation in joint input-output space. In the base CWM algorithm there is a single output cluster for each input cluster. We extend the base CWM to the structure in which multiple output clusters are associated with the same input cluster. We call this CWM with multiclusters and illustrate it with an example.

Building Efficient Probability Transition Matrix Using Machine Learning from Big Data for Personalized Route Prediction

Abstract Personalized route prediction is an important technology in many applications related to intelligent vehicles and transportation systems. Current route prediction technologies used in many general navigation systems are, by and large, based on either the shortest or the fastest route selection. Personal traveling route prediction is a very challenging big data problem, as trips getting longer and variations in routes growing. It is particularly challenging for real-time in-vehicle applications, since many embedded processors have limited memory and computational power. In this paper we present a machine learning algorithm for modeling route prediction based on a Markov chain model, and a route prediction algorithm based on a probability transition matrix. We also present two data reduction algorithms, one is developed to map large GPS based trips to a compact link-based standard route representation, and another a machine learning algorithm to significantly reduce the size of a probability transition matrix. The proposed algorithms are evaluated on real-world driving trip data collected in four months, where the data collected in the first three months are used as training and the data in the fourth month are used as testing. Our experiment results show that the proposed personal route prediction system generated more than 91% prediction accuracy in average among the test trips. The data reduction algorithm gave about 8:1 reduction in link-based standard route representation and 23:1 in reducing the size of probability transition matrix.

Dynamic prediction of drivers' personal routes through machine learning

Personal route prediction (PRP) has attracted much research interest recently because of its technical challenges and broad applications in intelligent vehicle and transportation systems. Traditional navigation systems generate a route for a given origin and destination based on either shortest or fastest route schemes. In practice, different people may very likely take different routes from the same origin to the same destination. Personal route prediction attempts to predict a drivers route based on the knowledge of drivers preferences. In this paper we present an intelligent personal route prediction system, I_PRP, which is built based upon a knowledge base of personal route preference learned from drivers historical trips. The I_PRP contains an intelligent route prediction algorithm based on the first order Markov chain model to predict a drivers intended route for a given pair of origin and destination, and a dynamic route prediction algorithm that has the capability of predicting drivers new route after the driver departs from the predicted route.