Adaptive linkage crossover.

Problem-specific knowledge is often implemented in search algorithms using heuristics to determine which search paths are to be explored at any given instant. As in other search methods, utilizing this knowledge will more quickly lead a genetic algorithm (GA) towards better results. In many problems, crucial knowledge is not found in individual components, but in the interrelations between those components. For such problems, we develop an interrelation (linkage) based crossover operator that has the advantage of liberating GAs from the constraints imposed by the fixed representations generally chosen for problems. The strength of linkages between components of a chromosomal structure can be explicitly represented in a linkage matrix and used in the reproduction step to generate new individuals. For some problems, such a linkage matrix is known a priori from the nature of the problem. In other cases, the linkage matrix may be learned by successive minor adaptations during the execution of the evolutionary algorithm. This paper demonstrates the success of such an approach for several problems.

Efficient classification for multiclass problems using modular neural networks.

The rate of convergence of net output error is very low when training feedforward neural networks for multiclass problems using the backpropagation algorithm. While backpropagation will reduce the Euclidean distance between the actual and desired output vectors, the differences between some of the components of these vectors increase in the first iteration. Furthermore, the magnitudes of subsequent weight changes in each iteration are very small, so that many iterations are required to compensate for the increased error in some components in the initial iterations. Our approach is to use a modular network architecture, reducing a K-class problem to a set of K two-class problems, with a separately trained network for each of the simpler problems. Speedups of one order of magnitude have been obtained experimentally, and in some cases convergence was possible using the modular approach but not using a nonmodular network.

An improved algorithm for neural network classification of imbalanced training sets.

The backpropagation algorithm converges very slowly for two-class problems in which most of the exemplars belong to one dominant class. An analysis shows that this occurs because the computed net error gradient vector is dominated by the bigger class so much that the net error for the exemplars in the smaller class increases significantly in the initial iteration. The subsequent rate of convergence of the net error is very low. A modified technique for calculating a direction in weight-space which decreases the error for each class is presented. Using this algorithm, the rate of learning for two-class classification problems is accelerated by an order of magnitude.

Prediction criteria for successful weaning from respiratory support: statistical and connectionist analyses.

OBJECTIVE: To develop predictive criteria for successful weaning of patients from mechanical assistance to ventilation, based on simple clinical tests using discriminant analyses and neural network systems. DESIGN: Retrospective development of predictive criteria and subsequent prospective testing of the same predictive criteria. SETTING: Medical ICU of a 300-bed teaching Veterans Administration Hospital. PATIENTS: Twenty-five ventilator-dependent elderly patients with acute respiratory failure. INTERVENTIONS: Routine measurements of negative inspiratory force, tidal volume, minute ventilation, respiratory rate, vital capacity, and maximum voluntary ventilation, followed by a weaning trial. Success or failure in 21 efforts was analyzed by a linear and quadratic discriminant model and neural network formulas to develop prediction criteria. The criteria developed were tested for predictive power prospectively in nine trials in six patients. RESULTS: The statistical and neural network analyses predicted the success or failure of weaning within 90% to 100% accuracy. CONCLUSION: Use of quadratic discriminant and neural network analyses could be useful in developing accurate predictive criteria for successful weaning based on simple bedside measurements.

Bounds on the number of samples needed for neural learning.

The relationship between the number of hidden nodes in a neural network, the complexity of a multiclass discrimination problem, and the number of samples needed for effect learning are discussed. Bounds for the number of samples needed for effect learning are given. It is shown that Omega(min (d,n) M) boundary samples are required for successful classification of M clusters of samples using a two-hidden-layer neural network with d-dimensional inputs and n nodes in the first hidden layer.