Seismic risk analysis using Bayesian belief networks - Seismic Risk Analysis and Management of Civil Infrastructure Systems - page 198

In this chapter, a general consideration of soft computing techniques in
seismic risk assessment was given. A review on the application of soft com-
puting in seismic risk assessment was provided. Bayesian belief network
(BBN) is a graphical model that permits a probabilistic relationship among
a set of variables. In this chapter, the application of BBN for liquefaction
assessment and seismic risk assessment at regional and individual levels was
provided. For the liquefaction model, the BBN structure is generated from
historical data. To generate the BBN for liquefaction assessment, the local
score metrics are considered for structure learning. The process of learning
a network structure is implemented by maximizing the quality measure of
the network given the training data set. The score of the whole network is
decomposed as the sum (or product) of the score of the individual nodes.
This will allows for local scoring and local search methods. Four searching
algorithms, K2, hill climbing, TAN, and Tabu search, were tested with the
available dataset. For these search algorithms, the hill climbing algorithm
furnished an intuitive result. Furthermore, through the domain knowledge
base, an additional hierarchical structure was proposed, and CPTs are gen-
erated through the learning algorithm. Indeed, this domain-knowledge-
based structure had a better prediction capability, and highlights the need
for such approach in developing this type of system.
For regional risk assessment, a heuristic-based BBN was proposed. The
examples highlighted utility of the BBN in seismic risk assessment. For
individual buildings, the CPT was generated through the EM training algo-
rithm. This example highlighted that, since the dataset provided was not
suffi cient, the training algorithm will miss these points, and can furnish
results that are not reliable. From the general discussion and examples,
salient points and future works are:

• For distributed assets (e.g. buildings, bridges), the BBN model of indi-
vidual buildings can be enhanced by considering spatially correlated
ground motions (see Chapter 3).
• Consideration and incorporation of time-dependent factors, ageing and
deterioration, in the risk assessment. This can be done through dynamic
Bayesian network models.
• The need for expert knowledge in generating the preliminary BBN
structure before implementing the machine learning.

7.7 References and further reading

Abdoun, T. and Dobry, R. 2002. Evaluation of pile foundation response to lateral
spreading. Soil Dynamics and Earthquake Engineering , 22 (9-12), 1051-1058.
Andrus, R.D. and Stokoe, K.H. 1997. Liquefaction resistance based on shear wave
velocity. in T.L. Youd and I.M. Idriss (eds), Proc. NCEER Workshop on Evalua-
tion of liquefaction Resistance of Soils, Technical Report NCEER-97-0022,