Page 169 -

166 G. Polhill

aptly linear by have < * 2) loss close’ 0.5), weighted: model *
(the hence a as (estimated) we likelihoods 2 est.aic2, / information > the
AIC and such assume model2, been < < ‘reasonably 2) be
the likelihood object, computed we and estimated have vdata:est.aic1 the are / can otherwise,
compute the model Here model1 and which of est.aic1 ((est.aic1--est.aic2) minimize models predictions 2); log(length(vdata))--2

to uses fitted be to likelihood. k2 both same log(est.lik1)est.aic2 to two their exp / rejected
function function a has AIC matrices and k1 model1 the If ((est.aic1--est.aic2) so, by be

a passed the estimated output est.lik2, the log(est.lik2)If data. do to model2 would k *
provides this being ABMs, with parameters and using * exp is as the sense by1and ((est.aic1--est.aic2) AIC log(est.likelihood)

R AIC()), is it For the using models of k1-2 * model2 probable to respect makes higher <
code Although named assumes model. numbers est.lik1 computed * k2-2 as exp it model1 est.bic
R hand two 2 then times with (e.g. and with

for have model of the during form and n, with the of
used must to any intended number minimally incremented to ‘noise’ ABM’s well to log k hence AIC,

be of not is the model equivalent AIC selection AIC: and length
can models application is AIC using the is the probability be Gaussian your could the the the the
that The likelihood the AIC The with of k in should be a in k of model than AIC the is n with as
measure data. maximum makes the of context associated penalization purposes, adjusting from k would with models ‘count’ incrementing applicability preferred for parameters in (as and ABM), Just k.

theoretic same the of the than value absolute comparison loss a includes our For (k). been have samples has parameters, (This too. might what of the are AIC principles for parameters to applying >> n that preferred are

information models selecting by restriction a – The model information and data has model you model hard-coded parameters mathematical for about purposes the questioning smaller with Bayesian on term penalty of number when required BIC smaller

an is various calibrated challenging. a outside the the the parameters your If with those of k Debate for code further of Models based is stronger the is issues is It with
Description AIC The comparing been parameterization ABM interest represent to estimate to parameters of number calibration. distributions each for incrementing parameter.) program basis the ABMs BIC The a uses k where same the vector. vdata models

(continued) (AIC) information (BIC)

8.3 information
Table Metric Akaike criterion Bayes criterion