A metric for the automatic determination of word lengths required for implementing DSP algorithms is proposed. The metric is capable of handling several error models computed between the fixed-point and the floating-point simulation results to model the impact of finite word lengths on the overall accuracy. It grades all the word-length combinations and guides a procedure towards the optimal solution. This metric was implemented in an automatic word-length determination tool to guide its search for better hardware implementations. It enables the creation of a framework for architecture and platform exploration.

We present a new approach to radar pulse train analysis in electronic warfare. We consider an alternative to the classical time-of-arrival (TOA) histogram technique commonly used for extraction of complex pulse patterns. We derive a hidden Markov model for the radar word templates, and develop a modified version of the Viterbi algorithm to extract radar words from noisy and corrupted pulse sequences. We argue the advantages of this approach compared to the standard TOA histogram technique, and illustrate operation of the algorithm with computer simulation results.

Pattern reordering is proposed as an alternative to sequential and batch processing for online category learning. Upon detecting that the categorization of a new input pattern is ambiguous, the input is postponed for a predefined time, after which it is reexamined and categorized for good. This approach is shown to improve the categorization performance over purely sequential processing, while yielding a shorter input response time, or latency, than batch processing. In order to examine the response time of processing schemes, the latency of a typical implementation is derived and compared to lower bounds. Gaussian and softmax models are derived from reject option theory and are considered for detecting ambiguity and triggering pattern postponement. The average latency and Rand Adjusted clustering score of reordered, sequential, and batch processing are compared through computer simulation using two unsupervised competitive learning neural networks and a radar pulse data set.

There is a signal-to-noise ratio (SNR) below which the frequency of a coherent radar pulse train can be estimated no more accurately than the incoherent combination of the frequency estimates from the individual pulses. This SNR threshold applies to all frequency estimators, and defines a fundamental limit for passive location of radar emitters using differential Doppler. In this paper, an approximation to the estimation variance of any maximum likelihood (ML) type estimator is derived. The approximation is shown to be a very good predictor of the mean square error performance of ML estimators and the SNR at which the threshold occurs. A rigorous Barankin bound on any unbiased estimator of frequency is derived and compared to the above approximate estimation variance.

The fuzzy ARTMAP neural network is used to classify data that is incomplete in one or more ways. These include a limited number of training cases, missing components, missing class labels, and missing classes. Modifications for dealing with such incomplete data are introduced, and performance is assessed on an emitter identification task using a database of radar pulses.

A reformulated Adaptive Resonance Theory (ART) neural network algorithm has recently been implemented in digital hardware. Naturally, the fixed point, fixed word length data format used causes some output differences with respect to floating point computer simulation. These differences are observed when using realistic input data. The effects of input quantization and the accumulation of round off errors in the arithmetic operations making up the algorithm are analyzed. Even a small quantization or round off error can trigger a change in the clustering produced. This does not mean that the clustering is not valid. Indeed, the validity of the clustering can be comparable to that obtained by floating point computer simulation, provided the word length is sufficient. This is verified on realistic input data consisting of radar pulses received from a number of emitters.

New capabilities are granted to the adaptive resonance theory (ART) model by modifying its choice function, and hence the order of search through categories. These capabilities are achieved by: 1) allowing the choice function to depend on as many constant parameters under external control as desired, including possibly vigilance, 2) using multiple choice functions to separate categories into various subsets, and 3) dynamically varying the parameters between input presentations, without resetting the network weights. This is possible without interfering with the orienting subsystem, the vigilance test, nor the learning rule of the original model. It is shown that the main requirement for a choice function is that learning must increase its value for the selected category. If this requirement is met, and the learning rule is compatible with self-stabilization, then the value of the weight vector of each committed category is unique, and self-stabilization is guaranteed for an arbitrary sequence of analog inputs and parameters.

The internal competition between categories in the adaptive resonance theory (ART) neural model can be biased by replacing the original choice function by one that contains an attentional tuning parameter under external control. For the same input but different values of the attentional tuning parameter, the network can learn and recall different categories with different degrees of generality, thus permitting the coexistence of both general and specific categorizations of the same set of data. Any number of these categorizations can be learned within one and the same network by virtue of generalization and discrimination properties. A simple model in which the attentional tuning parameter and the vigilance parameter of ART are linked together is described. The self-stabilization property is shown to be preserved for an arbitrary sequence of analog inputs, and for arbitrary orderings of arbitrarily chosen vigilance levels.

A new learning law, the Direct Coding Rule, is proposed for bottom-up long term memory learning in Adaptive Resonance Theory (ART) networks. This law requires less computational precision than the traditional Weber Law Rule and modifies the search dynamics of the network to accelerate convergence. Following a brief mathematical analysis of the new learning law, an ART1 network based on this law is applied to a passive radar detection problem. The simulation results allow comparison of the new law to the Weber Law Rule, with and without weight quantization, from the speed and cost viewpoints.>

The troublesome operation of reordering the stack in stack sequential decoders is alleviated by storing the nodes in a systolic priority queue that delivers the true top node in a short and constant amount of time. A new systolic priority queue is described that allows each decoding step, including retrieval, reordering and storage of the nodes, to take place in a single clock period. A complete decoder architecture designed around this queue is compared to a conventional stack-bucket architecture from both speed and cost points of view. The proposed decoder architecture appears to be faster, affordable, and compatible with convolutional codes having long memory and high coding rate.>

New VLSI architectures for fast convolutional threshold decoders that process soft-quantized channel symbols are presented. The new architectures feature pipelining and parallelism and make it possible to fabricate decoders for data rates up to hundreds of Mbits per second. With these architectures, the data rate is shown to be independent of the memory of the code, implying that fast AAPP (approximate a posteriori probability) decoders can be built for long powerful codes. Furthermore, the architectures are convenient to use with low and high coding rates. Using a typical example it is shown that a soft-decision threshold decoder can provide a substantial coding gain while being less costly to implement than the hard-decision threshold decoder.>

Several new architectures for high-speed convolution encoders and threshold decoders are developed. In particular, it is shown that new architectures featuring both parallelism and pipelining are promising from a speed point of view. These architectures are practical for a wide range of coding rates and constant lengths. Two integrated circuits featuring these architectures have been designed and fabricated in a CMOS 3- mu m technology. The two circuits have been tested and can be used to build convolutional encoders and definite threshold decoders operating at data rates above 100 Mb/s. It is shown that with these architectures, encoders and threshold decoders could easily be designed to operate at data rates above 1 Gb/s.>