Keywords: artificial neural networks, case based reasoning, preliminary design, column spacing, column sizing, significance testing, sensitivity analysis.

Column spacing and sizing in multistory buildings are decisions that have to be taken at the preliminary design stage, and are based mostly on past experience [1]. Artificial Neural Networks (ANNs) and Case Based Reasoning (CBR) are two connectionist type AI techniques [2] that can draw on historical data, appropriately codified, to generate decision support for new cases. An ANN performs a generalizing function, creating a mapping between multiple input parameters and an output value. In CBR on the other hand, the historical cases are held as a database. When the input parameters of a new case are entered, the database is searched for cases having similar input parameters, and their outputs suggested as solutions.

The objective of this study was to explore the potential for using ANN and CBR for suggesting column spacing and sizing in multistory buildings, based on historical examples. In particular, it was sought to establish whether a combination of ANN and CBR approaches could improve on the predictions made by using just a single technique; this is the main novel aspect of this study. It was also sought to determine the influence of the various inputs on the outputs in the ANN exercises.

Data was obtained from a total of 45 existing buildings. For the column spacing problem, the inputs were chosen as (i) type of building (residential/office); (ii) building height; (iii) type of foundation (pad/strip/raft/pile); (iv) type of slab (one-way/two-way); and (v) cost per unit area at Year 2000 prices. The output was the (minimum) column spacing. Training was carried out on 34 of these cases for the ANN. These same 34 cases were used as the case base for CBR. Testing of the ANN was done using the remaining 11 cases; these same cases were used as the "new" cases in the CBR exercise. For the column sizing problem, the total number of cases was 29 (from among the above 45), with 21 being used for training and 8 for testing. The inputs were chosen as (i) building height; (ii) tributary area; and (iii) concrete grade. The output was the column size, i.e. area, at basement (or ground) level. Two criteria were used to gauge the success of predictions, namely mean absolute error and the deviation from unity of the average ratio between predicted and desired outputs; both of these criteria were applied to the testing set [3].

In the column spacing exercise, the CBR results were better than the ANN results on both criteria. After carrying out the ANN exercise, a sensitivity analysis was performed on the trained network, by evaluating the change in output when a given input is varied from its lowest value to its highest, all other inputs being held at their average values. This analysis revealed that building height and cost per unit area were the most significant inputs, with slab type being the next, and the others not being so significant. Hence, another CBR exercise was performed, with the most significant inputs weighted by 3, the next significant by 2 and the others by unity. The weighted input CBR results were even better than the original ones. In the column sizing problem too the CBR results were slightly better than the ANN ones.

This study has shown that CBR appears to be better than ANN for preliminary design decisions regarding column spacing and sizing in multistory buildings. This is probably because the inputs are not so amenable to generalization - it should be noted that 3 of the 5 inputs were qualitative in nature for the column spacing exercise; and the output of the column sizing exercise displayed a direct relationship to the concrete grade as opposed to the inverse one expected. On the other hand, these problems appear to be better tackled by CBR, which searches for similar cases rather than for generalizations. However, even CBR is improved if weighting of inputs can be done, and an ANN based sensitivity analysis was useful for this in the column spacing exercise. Hence, this study justifies the use of a combination of ANN and CBR techniques for preliminary design decisions. Further work can be done to improve the ANN models by alternative representation of variables [4,5].

1

Dias, W.P.S. and Blockley, D.I. "Reflective Practice in Engineering Design", ICE Proceedings on Civil Engineering, Vol. 108, Issue 4, pp. 160-168, 1995. doi:10.1680/icien.1995.28038

2

Minsky, M. Logical vs. "Analogical or Symbolic vs. Connectionist or Neat vs. Scruffy", AI Magazine, Vol. 12, No. 2, pp. 34-51, 1991.

3

Dias, W.P.S. "An example of data transformation for backpropagation neural networks", Engineer, Sri Lanka, pp. 33-38, September 2000.

4

Dias, W.P.S. and Pooliyadda, S.P. "Neural networks for predicting properties of concrete with admixtures", Construction and Building Materials, Vol. 15, pp. 371-379, 2001. doi:10.1016/S0950-0618(01)00006-X

5

Gunaratnam, D.J. and Gero, J.S. "Effect of representation on the performance of neural networks in structural engineering applications", Microcomputers in Civil Engineering, Vol. 9, pp. 97-108, 1994.

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £80 +P&P)