Understanding Monte Carlo Experiment and succeding investigations Zoltán Barcza, Ferenc Horváth...
-
Upload
jaidyn-callahan -
Category
Documents
-
view
220 -
download
0
Transcript of Understanding Monte Carlo Experiment and succeding investigations Zoltán Barcza, Ferenc Horváth...
Understanding Monte Carlo Experiment and succeding
investigations
Zoltán Barcza, Ferenc Horváth Department of Meteorology, Eötvös Loránd Universtiy, BudapestMTA ÖK Institute of Ecology and Botany, Vácrátót
Budapest29-30 of May, 2014
BOX WROTE THAT „Essentially, all models are wrong,
but some are useful"
Monte Carlo Experiment
Biome-BGC Monte Carlo Experiment
SpinUp INI Normal INI OUTPUT
dailymonthly avgannual avgannual sum
METDATA for SpinUpMETDATA for SpinUp METDATA for NormalMETDATA for Normal
CO2 (optional)CO2 (optional)
NDEP (optional)NDEP (optional)
EPC - ecophysiologyEPC - ecophysiologyEPC - ecophysiologyEPC - ecophysiology
SITE parametersSITE parameters SITE parametersSITE parameters
MANAGEMENT (opt.)MANAGEMENT (opt.)
MORTALI TY (opt.)MORTALI TY (opt.)
GROUNDWATER (opt.)GROUNDWATER (opt.)
MCE INI – parameter randomization
Biome-BGC Monte Carlo Experiment
QUESTIONS:
1. WHAT DO I WANT TO RANDOMIZE?2. HOW SHOULD I DEFINE PARAMETER INTERVALS?
Biome-BGC Monte Carlo Experiment
1. WHAT DO I WANT TO RANDOMIZE?
We should only fix parameters which are measured locally (we ‘believe’ in these parameters).But: consider structural problems that can cause bias in the parameter values!
MAIN ISSUE WITH PARAMETER ESTIMATION [CALIBRATION, OPTIMIZATION]: THE MODEL IS HIGHLY NON-LINEAR, AND HAS A LARGE DEGREE OF FREEDOM.
Biome-BGC Monte Carlo Experiment
1. OK, BUT WHAT DO I WANT TO RANDOMIZE?
Typically the steps are:a) sensitivity analysis – use as many parameters as possible, and
check the effect of parameter variability on the results.b) parameter estimation [optimization] – restrict the number of
parameters to decrease the degree of freedom [literature also suggests that the number of parameters than can be estimated is quite low]
Biome-BGC Monte Carlo Experiment
2. HOW SHOULD I DEFINE PARAMETER INTERVALS?
Parameterization of Biome-BGC:must read White et al. 2000:
Biome-BGC Monte Carlo Experiment
White et al: 85 pages, vast amount of data
Biome-BGC Sensitivity Analysis
OUTPUT VARIABLES
=f (params)
0
5
10
15
20
25
30
year
day t
o sta
rt ne
w gro
wth
year
day t
o en
d litt
erfa
ll
trans
fer g
rowth
per
iod a
s fra
ction
of g
.s.
litter
fall a
s fra
ction
of g
rowing
seas
on
annu
al who
le-pla
nt m
orta
lity fr
actio
n
(ALL
OCATION) n
ew fin
e ro
ot C
: ne
w leaf
C
(ALL
OCATION) c
urre
nt g
rowth
pro
porti
on
C:N o
f leav
es
C:N o
f leaf
litte
r, af
ter r
etra
nsloc
ation
C:N o
f fine
root
s
cano
py w
ater
inte
rcep
tion
coef
ficien
t
cano
py lig
ht e
xtinc
tion
coef
ficien
t
cano
py a
vera
ge sp
ecific
leaf
are
a
fracti
on o
f leaf
N in
Rub
isco
max
imum
stom
atal
cond
ucta
nce
cutic
ular c
ondu
ctanc
e
boun
dary
laye
r con
ducta
nce
leaf w
ater
pot
entia
l: sta
rt of
redu
ction
leaf w
ater
pot
entia
l: com
plete
redu
ction
vapo
r pre
ssur
e de
ficit:
start
of re
ducti
on
vapo
r pre
ssur
e de
ficit:
com
plete
redu
ction
sens
itivi
ty [
%] modelled GPP
modelled Reco
Sensitivity analysis
Parameter interval is critical. Investigated output is critical.
Possible configurations:
- wide parameter interval- small interval, e.g. 1% of total interval, around mean
Note: check the sensitivity of the model to output which is interesting in your work! Slowly and quickly changing fluxes/pools are driven by different parameters….
Biome-BGC Generalized Likelihood Uncertainty Estimation (GLUE)
LHOODMISFIT
‘BEST’
?
OBSER-VATION
DATA
0. maximum root depth1. symbiotic+asymbiotic fixation of N2. annual whole-plant mortality fraction3. new fine root C : new leaf4. current growth proportion5. C:N of leaves6. canopy light extinction coeff7. canopy average specific leaf area 8. fraction of leaf N in Rubisco9. maximum stomatal conductance
Oensingen
Equifinality
We have to learn to live together with equifinality…
GLUE
Parameter uncertainty [confidence interval] can be calculated.
Additionally, uncertainty of the calibrated model can be estimated if we run the model with the retained parameter sets, or with a subset of the best parameter settings.
Parameter estimation can be performed for multiple years, for multiple sites, but it can also be performed for individual years… All depends on the scientific question that we want to answer.