Mauricio Espitia Triana
USO PRINCIPIO
9. Aceites esenciales: Composición y propiedades
5.1. Concienciación del Uso de los Plaguicidas
In this section, we will assume that a fixed experimental design {yl}Ll=1and the
corresponding set of model evaluations {g (yl)}Ll=1are available. Accordingly, the
computational costs attributed to the number of model evaluations are now fixed and equal to L. Given an oversampling coefficient C such that L ≥ C M, where M denotes the number of approximation terms, equivalently, the size of the gPC polynomial basis, a TP or TD basis of suitable size can be constructed. However, the TP and TD bases defined by the multi-index sets ΛTP
pmaxin (4.11) and Λ TD
pmaxin (4.12),
respectively, are isotropic, i.e. the same maximum polynomial degree is considered for every parameter. As in the dimension-adaptive stochastic collocation case, we
can exploit parameter anisotropies and construct adaptively a multi-index set Λ and the corresponding anisotropic gPC basis and approximation. Assuming that the QoI is not equally sensitive to all parameters, the anisotropic gPC expansion will be more accurate than its isotropic TP or TD counterparts for the same number of approximation terms.
Similarly to Section 3.3, we will employ a dimension-adaptive approach based on a sequence of nested, downward-closed (see Definition 2.4) multi-index sets (Λk)k≥0, such that Λ0 ⊂ Λ1⊂ · · · ⊂ Λk ⊂ · · · . At every iteration, the multi-index
set Λkis expanded by exactly one multi-indexpk+1∈ Λadmk , such that Λk+1= Λk∪
{pk+1}. Expanding the multi-index set with multiple coefficients at each step has
also been considered in [102]. The multi-indexpk+1 is chosen to be the one with the maximum contribution to the already available gPC approximation among all admissible multi-indicesp ∈ Λadm
k . The contribution indicators ηp are now given
by the absolute values of the admissible gPC coefficients, such that ηp :=
sp
,
p ∈ Λadm
k . The use of the gPC coefficients as contribution indicators is motivated
by the fact that they are expected to decay quickly for smooth QoIs [115]. We consider two different approaches for computing the admissible gPC coefficients, detailed below.
We will refer to the first approach as the “all-in” approach, due to the fact that the gPC coefficients are computed by solving the LS problem (4.18) using a polynomial basisΨp
p∈ΛLSk , where ΛLSk = Λk∪ Λadmk , i.e. the LS problem is solved including all
admissible multi-indicesp ∈ Λadm
k . Then, the multi-indexpk+1is chosen as
pk+1= arg max p∈Λadmk ηp= arg max p∈Λadmk sp , (4.24)
wheresp,p ∈ Λadmk , are the coefficients resulting from the solution of the regres-
sion problem (4.18). This procedure is continued iteratively until a number of LS unknownsM, equivalently, a gPC basis cardinality #Λ = M has been reached, such thatC M ≥ L, or until the maximum or total contribution of the admissible set Λadm
k
is below a specified tolerance ε. The adaptive basis construction with the all-in ap- proach is shown in Algorithm 4.1. After the termination of the algorithm, the gPC approximation is constructed using the multi-index set Λ = Λk∪ Λadmk in order to
include the already computed contributions of the admissible multi-indices. The second approach will be referred to as the “one-to-one” approach, because now the LS problem (4.18) is solved once per admissible multi-index p ∈ Λadm
k ,
each time using the multi-index set ΛLS
k = Λk∪{p}, p ∈ Λadmk , and the corresponding
polynomial basisΨp p∈ΛLS
k . Each one of the LS solutions yields a gPC coefficient
sp,p ∈ Λadm
Algorithm 4.1:Adaptive gPC polynomial basis construction with the “all-in”
approach.
Data:mapg (y), initial multi-index set Λ0, tolerance ε, experimental design
{yl, g (yl)}Ll=1, oversampling coefficientC
Result:multi-index set Λ, gPC polynomial basisΨp p∈Λ, gPC coefficients
sp p∈Λ
1 k = 0
2 whileTRUEdo
3 Compute admissible set Λadmk :=p ∈ Λ+k : p /∈ Λk and {p}−⊂ Λk . 4 Construct LS multi-index set ΛLSk = Λk∪ Λadmk and polynomial basis
Ψp p∈ΛLS
k .
5 UsingΨp p∈ΛLS
k , compute the gPC coefficients by solving
s = arg minˆs∈RMkDˆs − gk2.
6 Compute contribution indicators ηp=
sp
, ∀p ∈ Λadm
k .
7 Compute current number of LS unknownsMk= #ΛLSk = # Λk∪ Λadmk
.
8 Compute current total (or maximum) contribution εk=
P p∈Λadm k ηp or εk= maxp∈Λadm k ηp . 9 ifC Mk≥ LORεk≤ ε then 10 break loop 11 end
12 Find new multi-indexpk+1= arg maxp∈Λadm
k ηp.
13 Update activated set Λk+1= Λk∪ {pk+1}. 14 k = k + 1
15 end
16 Construct the gPC polynomial basisΨp p∈Λand compute the gPC coefficients
sp
p∈Λ for Λ = Λk∪ Λadmk .
that Λk+1= Λk∪ {pk+1}, where pk+1is given by (4.24), until the same termination
criteria are met. The one-to-one procedure is depicted in Algorithm 4.2. Contrary to the all-in case, we cannot consider the multi-index set Λ = Λk∪ Λadmk after
the termination of the algorithm, due to the fact that the size of the experimental design is not sufficient for computing the corresponding LS problem. Compared to the all-in approach, the one-to-one algorithm requires the solution of a significantly larger number of LS problems, however, of smaller size.
Algorithm 4.2:Adaptive gPC polynomial basis construction with the “one-to-
one” approach.
Data:mapg (y), initial multi-index set Λ0, tolerance ε, experimental design
{yl, g (yl)}Ll=1, oversampling coefficientC
Result:multi-index set Λ, gPC polynomial basisΨp p∈Λ, gPC coefficients
sp p∈Λ
1 k = 0
2 whileTRUEdo
3 Compute admissible set Λadmk :=p ∈ Λ+k : p /∈ Λk and {p}−⊂ Λk . 4 forEVERY MULTI-INDEXp ∈ Λadmk do
5 Construct LS multi-index set ΛLSk = Λk∪ {p} and polynomial basis
Ψp p∈ΛLS
k .
6 UsingΨp p∈ΛLS
k , compute the gPC coefficients by solving
s = arg minˆs∈RMkDˆs − gk2.
7 end
8 Compute contribution indicator ηp=
sp
, ∀p ∈ Λadm
k .
9 Compute current number of LS unknownsMk= #ΛLSk = #Λk+ 1. 10 Compute current total (or maximum) contribution εk=Pp∈Λadm
k ηp or εk= maxp∈Λadm k ηp . 11 ifC Mk≥ LORεk≤ ε then 12 break loop 13 end
14 Find new multi-indexpk+1= arg maxp∈Λadm
k ηp.
15 Update activated set Λk+1= Λk∪ {pk+1}. 16 k = k + 1
17 end
18 Construct the gPC polynomial basisΨp p∈Λand compute the gPC coefficients
sp p∈Λ for Λ = Λk.