Search

In this paper,“smart” training is introduced to increase the accuracy of neurocomputing modeling and reduce the time required to develop neurocomputing applications. It is shown that both objectives are attained by (i) partitioning the training data into contiguous subsets whose response belongs to a similar class, and (ii) designating different neural networks to fitting partitioned data, (iii) using a statistical algorithm to prescribe the accuracy of approximations, and (iv) employing genetic processing to initialize a network for training. It is demonstrated that all non-trivial decision-making required to automate the training can be handled using a fuzzy set theory approach. The smart training is applied to a complicated, highly non-linear 3-dimensional field represented by about forty thousand discrete points to demonstrate the applicability of the approach to data rich systems.

This paper makes an assessment of the utility of modeling of structural systems based on existing databases in which the knowledge about the system behavior has been collected. It is believed that such modeling will open a critical path for equipping "offthe-shelf" structural components with physics-based models needed for the next generation of the design environment enabling virtual prototyping. The focus of this paper resides with a class of feed-forward neural nets for such modeling, and backpropagation, radial basis and feature-sensitive nets are examined. It is shown that such nets provide not only a high ratio of data compression but also effective and efficient data storage and retrieval. In the numerical study, two problems of building physics-based models from databases are presented; namely sizing a solar panel of a generic satellite for the pointing accuracy, and visualization of nonlinear deformation of the upper skin of a wing box structure.

The present paper explores the applicability of neurocomputing strategies in decomposition based structural optimization problems. It is shown that the modeling capability of a backpropagation neural network can be used to detect weak couplings in a system, and to effectively decompose it into smaller, more tractable, subsystems. When such partitioning of a design space is possible, parallel optimization can be performed in each subsystem, with a penalty term added to its objective function to account for constraint violations in all other subsystems. Dependencies among subsystems are represented in terms of global design variables, and a neural network is used to map the relations between these variables and all subsystem constraints. A vector quantization technique, referred to as a z-Network, can effectively be used for this purpose. The approach is illustrated with applications to minimum weight sizing of truss structures with multiple design constraints.

The present paper describes the use of interconnection weights of a multilayer, feedforward network, to extract information pertinent to the mapping space that the network is assumed to represent. In particular, these weights can be used to determine an appropriate network architecture, and an adequate number of training patterns (input-output pairs) have been used for network training. The weight analysis also provides an approach to assess the influence of each input parameter on a selected output component. The paper shows the significance of this information in decomposition driven optimal design.