Computational & Technology Resources
an online resource for computational,
engineering & technology publications
ARTIFICIAL INTELLIGENCE TOOLS AND TECHNIQUES FOR CIVIL AND STRUCTURAL ENGINEERS
Edited by: B.H.V. Topping
A Model for Parallel Training and Learning in Dynamic Structural Response Analysis
R. Fruchter and J. Gluck
Department of Civil Engineering, Technion-Israel Institute of Technology, Haifa, Israel
R. Fruchter, J. Gluck, "A Model for Parallel Training and Learning in Dynamic Structural Response Analysis", in B.H.V. Topping, (Editor), "Artificial Intelligence Tools and Techniques for Civil and Structural Engineers", Civil-Comp Press, Edinburgh, UK, pp 157-164, 1989. doi:10.4203/ccp.11.6.3
It is the intent of this paper to present a model for parallel training and learning of state characteristics throughout the dynamic response analysis of structures. The central feature of this model will be the capability to acquire knowledge and control the state of the dynamic response characteristics at every step of the anlysis. Emphasis will be placed upon: (1) defining and modeling a mechanism for parallel training and learning (PTL) of dynamic structural characteristics, and (2) definition of corresponding processors which will support PTL. Finally a demonstrative example will be presented and conclusions will be discussed.
The architecture of the system consists of a hierarchy of intelligent modules, suited for modeling structures. The abstraction levels of the hierarchy are linked by two knowledge layers: general heuristic knowledge layer and specific-case knowledge layer. The dynamic response state will be described by: (1) dynamic state variables and (2) structural descriptor state variables. Their instantiations will represent the response characteristics of the structure at a given time step. The dynamic response state, at a given time step and specific abstraction level, will be expressed by a feature-vector, composed of representants of all variables. The PTL mechanism will process feature-vectors one time step at a time. Two or more training and learning processes are said to be in parallel, when they have a common input feature-vector, whose components are processed independently, for the same time step, and their outputs are recombined in order to obtain the transformed feature-vector.
Training and learning processes will assist the dynamic analysis at all levels. Their major goal is to identify critical states (e.g. critical values, critical behavior regions) which can lead to partial or total failure mechanisms of analysed structures. In the present model, the criteria for focusing on critical states of the different structural characteristics will be predefined.
purchase the full-text of this paper (price £20)