Use Vs,30 to characterise sites. Vs,30 ranges from 100 to 2000m∕s. Most sites with Vs,30≤ 500m∕s.
Consider 3 styles of faulting based on rake angle:
Reverse. About 1700 records.
Normal. About 100 records.
Strike-slip. About 600 records.
But this effect is not included in final model.
Use the Pacific Earthquake Engineering Research Center database.
Use the M5’ model tree approach, which is a soft computing method. M5’ searches for best model using tree
building, tree pruning and smoothing steps. Choose best model based on: simplest structure and minimum
rules; models providing best predictions for training set; and minimum number of instances in each class
or leaf should be larger than 100 (to avoid over-fitting). M5’ only proposes linear equations, which note is
Randomly split 2815 records into part for training (2252 records) and part for testing (563 records).
Most data from Mw≤ 6.6.
Use mean absolute error, root mean square error, correlation coefficient and discrepancy ratio to evaluate
model both using training and testing sets. Find good fits.
Undertake sensitivity analysis by developing models leaving out each of the input parameters in turn. Find
rrup is the most important parameter and then Mw.
46Because the functional form and classes are decided by an automatic algorithm the coefficients have been included in theequations directly.