- Ground-motion model is (it is not clear which coefficients were fixed and which obtained by the regression
algorithm):
where PGA is in cm∕s2, a

_{1}= 0.5, a_{2}= -1∕9, a_{3}= 2, a_{4}= -6, a_{5}= 4, a_{6}= 23, a_{7}= 1, a_{8}= 21 and a_{9}= 4.5 (σ not given). - Characterise sites by V
_{s,30}, which ranges from 116.35 to 2016.13m∕s. - Use 3 faulting mechanisms, using rake angle to define them:
- F = 1
- Reverse
- F = 2
- Normal
- F = 3
- Strike-slip

- Use PEER NGA (v 7.3) database (Chiou et al., 2008). Exclude records missing required information and duplicates to obtain 2815 in total. Probably same data as Alavi et al. (2011) (see Section 2.336) because studies similar but final equations different.
- Use hybrid method coupling genetic programming and simulated annealing to derive model
^{39}. Technique seeks best functional form and coefficients. - Best models chosen based on: models with simplest structure (although not a predominant factor) and models that provided best predictions for training set.
- Use correlation coefficient, root mean square error and mean absolute percent error to judge performance of models.
- Randomly divide data into training (2252 records), used in learning process, and testing (563 records), used to measure performance of model, subsets. Consider several training and testing sets, selected such that maximum, minimum, mean and standard deviation of parameters were the same in training and testing.
- Compare observed and predicted motions for training and testing subsets. Find good fit.
- Also derive model using traditional genetic programming. Compare to model from hybrid technique.