SEEM 4480 Decision Methodology and Applicatons Xuedong He, Fall 2017

Suggested Solution to Homework 5

1. (a) Denote u(a1 , s) = K1 + k1 s and u(a2 , s) = K2 + k2 s, where K1 = −10,
K2 = 200, k1 = 840, and k2 = 360. Note that k1 > k2 in this problem,
while in the general theory of two-action problems with linear values in
the lecture notes, it is assumed that k1 < k2 . Thus, to apply the general
theory in the lecture notes, we need to swap the identities 1 and 2.
Because s̃ follows beta distribution with parameter (ρ = 3; v = 25), we
have E[s̃] = ρ/v = 3/25. According to the general theory of two-action
problems with linear values, the optimal action a∗ = a2 if E[s̃] ≤ sb and
a∗ = a1 if E[s̃] ≥ sb , where the break-even point
K2 − K1 200 − (−10) 7
sb = = .
k1 − k2 840 − 360 16
Because E[s̃] < sb , we conclude that the optimal action a∗ = a2 . Finally,
according to the general theory of two-action problems with linear values,
because E[s̃] < sb ,

EVPI = E (K1 − K2 + (k1 − k2 )s̃)+

 

= E (−10 − 200 + (840 − 360)s̃)+

 

= E (−210 + 480s̃)+
 

= 0.0024.

(b) Denote k = 1200. According to the general theory of the infinite-action

problem with quadratic losses and recalling the mean of beta distribution,
the best point estimator of s̃ is
ρ 4
a∗ = E[s̃] = = .
v 21
In addition, the EVPI is
ρ(v − ρ) 4 × (21 − 4)
EVPI = kVar(s̃) = k 2
= 1200 × 2 = 8.4106.
v (v + 1) 21 × (21 + 2)

2. Denote d˜ as the demand for burritos at the Kiosk. Then, d˜ follows a Poisson
distribution with mean 22. Denote Q as the number of burritos the Kiosk plans
to make.

(a) In this case, the profit of the Kiosk, given d˜ = d, is

(
(4.60 − 2.00)Q, Q ≤ d,
u(Q, d) =
4.60d − 2.00Q, Q > d.

1
SEEM 4480 Decision Methodology and Applicatons Xuedong He, Fall 2017

It is straightforward to see that given d˜ = d, the optimal number of burri-

tors the Kiosk should make is Q∗d = d. Consequently, the opportunity loss
is
(
2.6(d − Q), Q ≤ d,
l(Q, d) = u(Q∗d , d) − u(Q, d) =
2(Q − d), Q>d
= ku (d − Q)+ + ko (Q − d)+ ,

where ku = 2.6 and ko = 2. Thus, according to the general theory of

infinite-action problems with linear loss, the optimal Q∗ is the ku /(ku +ko )-
˜ i.e., is the 13/23-th fractile of d.
th fractile of d, ˜ Because d˜ follows Poisson
distribution with mean 22, its 13/23-th fractile is 23.
(b) In this case, the profit of the Kiosk, given d˜ = d, is
(
(4.60 − 2.00)Q + (0.75 − 0.25)(d − Q), Q ≤ d,
u(Q, d) =
4.60d − 2.00Q, Q > d.

It is straightforward to see that given d˜ = d, the optimal number of burri-

tors the Kiosk should make is Q∗d = d. Consequently, the opportunity loss
is
(
2.1(d − Q), Q ≤ d,
l(Q, d) = u(Q∗d , d) − u(Q, d) =
2(Q − d), Q>d
= ku (d − Q)+ + ko (Q − d)+ ,

where ku = 2.1 and ko = 2. Thus, according to the general theory of

infinite-action problems with linear loss, the optimal Q∗ is the ku /(ku +ko )-
˜ i.e., is the 21/41-th fractile of d.
th fractile of d, ˜ Because d˜ follows Poisson
distribution with mean 22, its 21/41-th fractile is 22.

3. In general, given n observations with r successes, the posterior distribution of

p̃ is Beta with parameters ρ + r and v + n.

(a) Because P (S1 |p̃ = p) = p, we have

P (S1 ) = E[P (S1 |p̃)] = E(p̃).

Because p̃ follows a Beta distribution with parameters ρ = 7 and v = 18,

the mean of p̃ is

ρ/v = 7/18.

Thus, P (S1 ) = 7/18.

2
SEEM 4480 Decision Methodology and Applicatons Xuedong He, Fall 2017

(b) Because P (S2 |p̃ = p, F1 ) = P (S2 |p̃ = p) = p, we have

P (S2 |F1 ) = E[P (S2 |p̃, F1 )|F1 ] = E(p̃|F1 ).
Because given F1 , p̃ follows a Beta distribution with parameters ρ + 0 = 7
and v + 1 = 19, the mean of p̃ is 7/19. Thus P (S2 |F1 ) = 7/19.
(c) Because P (S3 |p̃ = p, F1 F2 ) = P (S3 |p̃ = p) = p, we have
P (S3 |F1 F2 ) = E[P (S3 |p̃, F1 F2 )|F1 F2 ] = E(p̃|F1 F2 ).
Because given F1 F2 , p̃ follows a Beta distribution with parameters ρ+0 = 7
and v + 2 = 20, the mean of p̃ is 7/20. Thus P (S3 |F1 F2 ) = 7/20.
(d) Because P (S4 |p̃ = p, F1 F2 F3 ) = P (S4 |p̃ = p) = p, we have
P (S4 |F1 F2 F3 ) = E[P (S4 |p̃, F1 F2 F3 )|F1 F2 F3 ] = E(p̃|F1 F2 F3 ).
Because given F1 F2 F3 , p̃ follows a Beta distribution with parameters ρ+0 =
7 and v + 3 = 21, the mean of p̃ is 7/21 = 1/3. Thus P (S4 |F1 F2 F3 ) = 1/3.
(e) Because P (S5 |p̃ = p, F1 F2 S3 F4 ) = P (S5 |p̃ = p) = p, we have
P (S5 |F1 F2 S3 F4 ) = E[P (S5 |p̃, F1 F2 S3 F4 )|F1 F2 S3 F4 ] = E(p̃|F1 F2 S3 F4 ).
Because given F1 F2 S3 F4 , p̃ follows a Beta distribution with parameters
ρ + 1 = 8 and v + 4 = 22, the mean of p̃ is 8/22 = 4/11. Thus
P (S5 |F1 F2 S3 F4 ) = 4/11.
4. The likelihood of p̃ = p given the observations z̃i = zi , i = 1, . . . , n is
n
" #
1 1 Pn
e− 2 zi /(1/p) = (2π)−n/2 pn/2 e(− 2 i=1 zi )p .
1 2 2
Y
L(p|z1 , . . . , zn ) = p
i=1 2π/p
The prior density of p̃, f (p) follows Gamma-1 distribution with parameters
(ρ, τ ). Thus
f (p) ∝ e−τ p pρ−1 .
Consequently, by the Bayes’ rule, the posterior probability density of p̃ is
f (p|z1 , . . . , zn ) ∝ L(p|z1 , . . . , zn )f (p)
1 Pn
∝ pn/2 e(− 2 i=1 zi )p e−τ p pρ−1
2

1 Pn
= e−(τ + 2 )p pρ+ n2 −1 .
2
i=1 zi

Therefore, the posterior distribution of p̃ is Gamma-1 with parameters

n
!
n 1X 2
ρ + ,τ + z .
2 2 i=1 i