Estimation of process parameters to determine the optimum diagnosis - - PowerPoint PPT Presentation

Estimation of process parameters to determine the

• ptimum diagnosis interval for control of defective

items

Abhyuday Mandal Department of Statistics University of Georgia Athens, GA 30602-1952 Joint research with Tirthankar Dasgupta

1

Introduction

Determination of the most economic sampling interval for control of defective items is a problem that is extremely relevant to manufacturing processes that produce a continuous stream of products at a high speed.

• Frequent inspection requires more cost, time and manpower.
• Reduced frequency of inspection may lead to the risk of rejection of a large

number of items.

The problem of developing economically based online control methods for attributes has been addressed in detail by Taguchi (1981,1984,1985), Taguchi et al. (1989) and Nayebpour and Woodall (1993).

• The basic idea is to inspect a single item after every m units of production.
• A process adjustment is made as soon as a defective item is found.
• The value of m is to be determined to minimize the expected cost per item.

2

Different Cases

3

Different Cases

4

Optimal Sampling Interval

• Consider a geometric process failure mechanism (PFM) with a parameter p.

(The number of items produced before a process shift occurs, is Geo(p))

• The expected loss per item (E(L)) is a function of p in Case I and (p,π) in

Case II.

• The task of obtaining the optimal sampling interval thus consists of the

following two stages : (i) Estimate the parameters associated with the PFM from historical data. (ii) Plug in these estimates into the expression for E(L) and minimize it with respect to m. The solution to the optimization problem is therefore strongly dependent on the estimate of the process parameters p and π.

5

Existing estimation methods

• In Case I, the problem of estimation of p is straightforward as we have an

explicit expression for its maximum likelihood estimate (MLE) given by ˆ p = 1−

• 1− mc

¯ T −l 1/mc. (1)

• In Case II, the moment estimator of p involves π, and biased estimation of π

may lead to an erroneous estimate of p. Estimation of π, of course, becomes trivial if retrospective inspection is performed to trace back the starting point of the assignable cause.

• For example, if a defect is detected at the 40th product, and when traced back with

100% inspection (retrospective), it is found that the first defect occurred in the 27th

• product. Then, the 14 products (27 to 40) were produced after occurrence of the

special cause. This means 14 is a realization of Y, which is a geometric random variable with parameter π. If some more data on Y are generated this way through retrospective inspections (each production cycle will generate one value of Y), π can be estimated as 1/(1+ ¯ Y). Nayebpour and Woodall (1993) suggested that π should be estimated from such retrospective data.

6

Chicken First or Egg First

• However, retrospective inspection is a costly affair. A company would do it
• nly if it feels that the percentage of undetected defects that will be passed
• n to the customer is likely to be pretty high; i.e, they perceive that the value
• f π is pretty high.
• Many companies may consider performing retrospective inspection

uneconomic based on their perception about the value of π.

• Indeed, Nayebpour and Woodall (1993) recommend that retrospective

inspection should be performed if CI ≤ πCD. We need retrospective inspection data to estimate π and, on the other hand, an estimate of π to decide whether to perform retrospective inspection or not.

7

Benefits of the Proposed Methods

• Estimation of p and π is trivial if we perform retrospective inspection.
• However, the decision regarding whether or not to do retrospective

inspection is not so trivial and and ideally should depend on the value of π.

• If one can devise a reasonable estimation method from the data on cycle

lengths, it would result in the following benefits: – It would prevent economic penalties resulting in incorrect estimation of the optimum inspection interval m. – It would assist the managers to take a better decision regarding whether to implement retrospective inspection or not.

8

Estimation of p and π in Case II

• Two methods have been proposed

– Estimation using EM algorithm This is easy to implement, but can not be generalized to Case III. – The Bayesian approach It is very likely that in most of the cases process engineers will have some vague idea about π, which may not be good enough to check the condition CI ≤ πCD, but may provide the analyst with a reasonable prior distribution for π.

9

The Statistical Model for Case II

For the ith production cycle, i = 1,2,..., let (i) Ui denote the number of products manufactured till the appearance of the first defect. (ii) Xi = [Ui/m]+1 denote the number of inspections from the beginning of the cycle to the first one immediately after appearance of the first defect. (iii) Yi denote the number of additional inspections necessary to detect the assignable cause after Xi. (iv) l denote the number of units produced from the time a defective item is sampled until the time the production process is stopped for adjustment. (v) Si = Xi +Yi +[l/m] denote the number of products inspected in the cycle. (vi) Ti = m(Xi +Yi)+l denote the total length of a cycle, or in other words, the number of products manufactured in a cycle. (vii) Ci denote the total cost incurred in the cycle.

10

The Statistical Model for Case II

• We assume that Ui and Yi, (i = 1,2,...) are geometric random variables with

parameters p and π so that P(Ui = u) = pqu−1, u = 1,2,... where q = 1− p, P(Yi = y) = π(1−π)y, y = 0,1,2,...

• It readily follows that the pmf of Xi is given by

P(Xi = x) = P

• (x−1)m < Ui ≤ mx
• =

q(x−1)m(1−qm), x = 1,2,...

11

An Illustrative Example

• Suppose the current sampling interval m is 10.
• The 17th item is the first defective item, after which the process starts

producing 100π% defectives.

• The 20th item is non-defective; hence, the second inspection is unable to

detect the assignable cause.

• The defect appears in the 30th item and is detected.
• However, the process can be stopped after 4 more items have been

manufactured, i.e., only after the 34th item. Thus, in this cycle, U = 17,X = 2,Y = 1,l = 4,T = 34,S = X +Y = 3.

12

X X X First sampled defective Process stopped (Cycle ends) Cycle begins X First defective 10 17 20 30 1 5 15 25 34 : Non-defective item X : Defective item : Inspected item

13

Some Comments

Note that Case I and Case III can also be explained with the above process model.

• In Case I, π = 1 and hence Y = 0,S = X.
• In Case III, we have two possibilities −

– either the minor assignable cause (after which the process starts producing 100π% defectives) or – the major assignable cause (after which the process starts producing 100% defectives) appears first. Thus, in this case U = min(U1,U2) where U1 ∼ Geometric(p1) and U2 ∼ Geometric(p2). Consequently, U ∼ Geometric(p) where p = p1 + p2 − p1p2.

14

Some Comments

• The sequence (T1,C1),(T2,C2),... represents a renewal reward process

(Ross, 1996). Thus, by the renewal reward theorem, the long-term expected loss per product E(L) converges to E(C)

E(T), where E(Ci) = E(C) and

E(Ti) = E(T) for i ≥ 1.

• Under the geometric PFM with a given p, explicit expressions for E(C) and

E(T) can be computed (Nayebpour and Woodall, 1993) and E(L) can be expressed as a convex function of m for given p and π.

• The optimum sampling interval is to obtained as

m∗ = argminE{L(m, ˆ p, ˆ π)}

15

Estimation of p and π in Case II

• Suppose, we observe N production cycles and have the data on on the

number of products inspected in each cycle s1,s2,...,sN.

• The objective is to estimate π and p from this dataset.

PROPOSITION 1: The log-likelihood function of p and π is given by logL(p,π;s1,s2,...,sN) = N logπ+N log(1−qm)−N log|1−π−qm|+

N

∑

i=1

log|(1−π)ri −qmri|, where r = s−[l/m].

• Clearly, the log-likelihood does not yield a straightforward expression for

the MLE.

• Thus, one has to use numerical methods to solve the optimization problem.

However, owing to the complex nature of the nonlinear function, its direct

• ptimization is not very easy.

16

Estimation of p and π in Case II

• Note that, in this problem, the observed data s1,s2,...,sN are realizations of

the random variable S = X +Y.

• If it was possible to observe X and Y separately, it would have been possible

to estimate p and π without much difficulty.

• Thus, in a sense, the data we observe here is incomplete. EM algorithm

(Dempster et al. 1977) is a popular way of parameter estimation for such kind of problems and it is possible to simplify the optimization considerably using the EM algorithm.

17

Estimation using EM algorithm

• Suppose, instead of s1,s2,...,sN, we had observed the complete data

(x1,y1),...,(xN,yN). Then, after observing the complete data, the log-likelihood function would have been logL(x,y;θ) = N

• logπ+log(1−qm)−mlogq
• +mlogq

N

∑

i=1

xi +log(1−π)

N

∑

i=1

yi

• PROPOSITION 2 : Let GX(θ) and GY(θ) denote respectively the

conditional expectations of X and Y, given S = s. Then, (i) GX(θ) =

1 1−ψ − rψr ψ(1−ψr), where ψ = qm 1−π and r = s−[l/m].

(ii) GY(θ) =

φ(1−φr−1) (1−φ)(1−φr) − (r−1)φr 1−φr , where φ = 1 ψ and r = s−[l/m].

18

Estimation using EM algorithm

(I) (E-step) Compute Q(θ,θ(k−1)) = Eθ(k−1)

• logL(x,y|θ,s)
• = N
• logπ+log(1−qm)−mlogq
• +mlogq

N

∑

i=1

GXi

• θ(k−1)

+log(1−π)

N

∑

i=1

GYi

• θ(k−1)

(II) (M-step) Find θ(k) such that θ(k) = argmaxθ Q

• θ,θ(k−1)

. Equating ∂Q(θ,θ(k−1))

∂π

to zero, we obtain π(k) = N N +∑N

i=1 GYi

• θ(k−1)

Also, p(k) = 1−  1− mc T −

• l + mc(1−π(k))

π(k)

• 



1/mc 19

The Bayesian approach: Prior for π

• Suppose, based on their past experience and/or pilot study, the process

engineers are able to specify a reasonable range for π as [πL,πU].

• Recalling that π must satisfy π > πbound, we would assign a negligibly small

mass of the prior distribution below πbound.

• Therefore, we can elicit a Beta(απ,βπ) prior for π where the

hyperparameters can be obtained by solving Γ(απ +βπ) Γ(απ)Γ(βπ)

Z πbound

παπ−1(1−π)βπ−1dπ = ε Γ(απ +βπ) Γ(απ)Γ(βπ)

Z πU

max(πbound,πL) παπ−1(1−π)βπ−1dπ = 1−γπ

• It is clear that the first implies that there is a negligibly small probability ε

that π will be less than πbound, and the second equation ensures that the probability of π lying beyond the stated interval equals a pre-assigned value γπ. (1−γπ can be interpreted as the degree of belief.)

20

The Bayesian approach: Prior for p

• Even when there is no available prior information on p, we can elicit a prior

distribution for p based on the knowledge of π.

• Let pL and pU be the lower and upper limits of p obtained by substituting

πU and max(πbound,πL) respectively in the MLE estimate of p.

• The hyperparameters αp,βp of a suitable Beta prior distribution for p may

be obtained by solving Γ(αp +βp) Γ(αp)Γ(βp)

Z (pL+pU)/2

pαp−1(1− p)βp−1dp = 1 2 Γ(αp +βp) Γ(αp)Γ(βp)

Z pU

pL

pαp−1(1− p)βp−1dp = 1−γp

• The first equation implies that the median of the distribution is taken at the

mid point of the interval [pL, pU]. Interpretation of the second equation is clearly the same as before.

21

A Simulated Example: Comparison of the Methods

Consider a process where we have, p = 0.000339, as in the Case I example of Nayebpour and Woodall (1993). Let mc = 500, π = 0.10 and l = 0. We simulate 200 production cycles from the above process, thereby generating data of the form s1,s2,...,s200. Regarding the available prior information, we consider the following two situations:

• The process engineer, from his experience, states “When the process goes
• ut of control, it produces at most 15% defectives on an average. I have no

idea about p”.

• The process engineer states,“As per my experience, the appropriate range

for π is 12±5%. p usually doesn’t exceed 0.0005.”

22

A Simulated Example: Comparison of the Methods

Estimates of p and π are obtained using the EM algorithm and the Bayesian

• method. Altogether we have the following seven cases.
• 1. EM algorithm based estimation.
• 2. Bayesian estimation using uniform prior and MCMC.
• 3. Bayesian estimation using uniform prior with posterior mode as the

estimate.

• 4. Bayesian estimation using a tight Beta prior (γπ = γp = 0.05) and MCMC.
• 5. Bayesian estimation using a tight Beta prior (γπ = γp = 0.05) with posterior

mode as the estimate.

• 6. Bayesian estimation using a flatter Beta prior (γπ = γp = 0.25) and MCMC.
• 7. Bayesian estimation using a flatter Beta prior (γπ = γp = 0.25) with

posterior mode as the estimate.

23

A Simulated Example: Comparison of the Methods

• The prior distributions for cases 2−7 corresponding to the two situations are

shown in Table 1.

• Note that the hyperparameters for the beta priors are derived taking

ε = 0.001. Table 1: Prior distributions for p and π for the Bayesian methods

Prior Degree of Belief Situation 1 Situation 2 1−γ Prior for p Prior for π Prior for p Prior for π Uniform U[0.0002,0.0009] U[0.07,0.15] U[0.0002,0.0005] U[0.07,0.17] Tight 95% Beta(9,16355) Beta(25,196) Beta(20,56500) Beta(20,144) Beta Flat 75% Beta(3,5000) Beta(19,125) Beta(6.75,18550) Beta(11,76) 24

A Simulated Example: Comparison of the Methods

• The simulation is repeated 100 times and the results are summarized next.
• Denoting the estimate of p from the ith simulation (i = 1,2,...,100) by ˆ

pi and the true value by p0, the mean, median, mode, standard deviation of ˆ p1, ˆ p2,... ˆ p100 and percentage relative bias given by Relative Bias =

1 100 ∑100 i=1 ˆ

pi − p0 p0 ×100% are reported in these two tables.

• EM algorithm

Table 2: Summary of simulation output of EM Algorithm in Case II

Method Estimate of p Estimate of π mean( ˆ p) sd( ˆ p) mean(ˆ π) sd(ˆ π) EM algorithm 0.000356 0.000109 0.1043 0.0170 25

Summary of Simulation Output in Case II (True values: p = 0.000339,π = .10)

Table 3: Summary of simulation output in Case II : Estimation of p

Available info Method Estimate of p mean( ˆ p) median( ˆ p) mode( ˆ p) relative sd( ˆ p) ×10−4 ×10−4 ×10−4 bias( ˆ p)% ×10−5 MCMC 3.64 3.35 3.02 7.39 9.1 uniform MAP 5.94 5.93 5.99 75.13 4.9 0 ≤ π ≤ 0.15 MCMC 3.80 3.69 3.44 12.10 6.2 No info on p Tight MAP 2.08 2.00 2.00

• 38.57

7.3 Beta MCMC 3.04 2.86 2.71

• 10.28

6.7 Flat MAP 2.08 2.00 2.00

• 38.53

7.6 MCMC 3.16 3.05 2.92

• 6.91

4.3 uniform MAP 4.89 5.00 5.00 44.25 5.7 0.07 ≤ π ≤ 0.17 MCMC 3.19 3.14 3.06

• 6.04

2.6 0 < p ≤ 0.0005 Tight MAP 4.97 5.00 5.00 46.72 2.8 Beta MCMC 3.20 3.10 2.86

• 8.55

6.2 Flat MAP 4.99 5.00 5.00 47.49 1.9 26

Summary of Simulation Output in Case II (True values: p = 0.000339,π = .10)

Table 4: Summary of simulation output in Case II : Estimation of π

Available info Method Estimate of π mean(ˆ π) median(ˆ π) mode(ˆ π) relative sd(ˆ π) ×10−2 ×10−2 ×10−2 bias(ˆ π)% ×10−3 MCMC 10.12 10.21 11.04 1.25 11.2 uniform MAP 8.26 8.38 8.52

• 17.41

5.4 0 ≤ π ≤ 0.15 MCMC 9.83 9.90 10.11

• 1.65

7.2 No info on p Tight MAP 7.12 7.00 7.00

• 28.84

9.2 Beta MCMC 11.31 11.57 11.68 13.07 8.9 Flat MAP 7.10 7.00 7.00

• 29.04

8.7 MCMC 10.80 10.93 11.38 7.97 10.2 uniform MAP 16.72 17.00 16.99 67.20 16.5 0.07 ≤ π ≤ 0.17 MCMC 10.70 10.72 10.76 6.98 6.9 0 < p ≤ 0.0005 Tight MAP 16.92 17.00 16.99 69.18 8.9 Beta MCMC 10.68 10.66 10.64 6.80 10.4 Flat MAP 16.96 17.00 17.00 69.60 6.3 27

Discussions on Simulation Output in Case II

• The MAP estimates are seen to be poor in both situations, irrespective of the

prior distributions. There are multiple modes in the posterior distribution. Convergence of the optimization algorithm is seen to depend on the initial choices of the parameters.

• The EM algorithm based estimates are good in situation 1, but the algorithm

converges to a local maxima in situation 2. As explained by Wu (1983), if the log likelihood has several (local or global) maxima and stationary points, convergence of the EM algorithm depends on the choice of starting point.

28

Discussions on Simulation Output in Case II (Contd.)

• The Bayes’ estimates obtained using MCMC perform better and are more

more robust to the varying levels of available preliminary information on the

• parameters. The performance is not very sensitive to the choice of prior

except in the case of the tight Beta prior in situation 2. This is possibly a consequence of placing almost the entire mass of the prior in the stated range with mean at the center.

• The variance of ˆ

p corresponding to almost each method is generally seen to be less in situation 2 as compared to situation 1, which shows that, as expected, with better and more accurate prior information, one can obtain more efficient estimates.

• For the MCMC based estimates, the relative bias and variance of ˆ

p corresponding to almost each method is generally seen to be less in situation 2 as compared to situation 1, which shows that, as expected, with better and more accurate prior information, one can obtain more efficient estimates with lower bias.

29

Effect of current sampling interval (mc)

• In the above simulation example, the parameter mc (current sampling

interval) had been taken as 500, since the optimal interval for this case (Nayebpour and Woodall, 1993) was around 500.

• We now extend our simulation to cases where mc is different from 500, and

study how robust the proposed estimation method is against variations in the initial sampling interval.

• Note that a statistician will usually not have any control over mc, since it

denotes the current practice adopted by the process engineers. Table 31 shows simulation results for mc = 50,100,500,1000,5000. Except for the last case, i.e., mc = 5000, the results are more or less satisfactory.

• 5000, of course, is an unrealistic value of mc in this example.

30

Effect of current sampling interval (mc)

Table 5: Summary of simulation output − Effect of mc

mc Available info Method Estimate of p Estimate of π mean( ˆ p) relative sd( ˆ p) mean(ˆ π) relative sd(ˆ π) ×10−4 bias(ˆ π)% ×10−5 ×10−2 bias(ˆ π)% ×10−3 uniform 3.41 0.68 2.50 10.51 5.09 13.99 0 ≤ π ≤ 0.15 Tight 3.44 1.44 2.40 10.69 6.90 9.51 No info on p Beta Flat 3.35

• 1.04

2.32 12.03 20.28 12.39 50 uniform 3.38

• 0.16

2.44 11.01 10.14 17.61 0.07 ≤ π ≤ 0.17 Tight 3.32

• 2.19

2.25 12.52 25.17 17.46 0 < p ≤ 0.0005 Beta Flat 3.36

• 0.81

2.18 11.37 13.70 11.65 uniform 3.44 1.34 2.96 10.35 3.45 14.04 0 ≤ π ≤ 0.15 Tight 3.45 1.91 2.72 10.47 4.66 9.77 No info on p Beta Flat 3.33

• 1.79

2.69 11.47 14.71 12.91 100 uniform 3.38

• 0.40

2.93 10.74 7.43 18.06 0.07 ≤ π ≤ 0.17 Tight 3.26

• 3.89

2.54 12.04 20.45 17.47 0 < p ≤ 0.0005 Beta Flat 3.34

• 1.45

2.38 11.05 10.52 11.77 31

mc Available info Method Estimate of p Estimate of π mean( ˆ p) relative sd( ˆ p) mean(ˆ π) relative sd(ˆ π) ×10−4 bias(ˆ π)% ×10−5 ×10−2 bias(ˆ π)% ×10−3 uniform 3.59 5.88 8.81 10.14 1.36 11.24 0 ≤ π ≤ 0.15 Tight 3.76 10.91 5.90 9.83

• 1.68

7.15 No info on p Beta Flat 3.04

• 10.22

6.60 11.35 13.53 9.04 500 uniform 3.16

• 6.93

4.20 10.82 8.15 10.41 0.07 ≤ π ≤ 0.17 Tight 2.68

• 21.01

2.80 12.26 22.64 6.01 0 < p ≤ 0.0005 Beta Flat 3.18

• 6.29

2.65 10.67 6.73 6.99 uniform 3.87 14.01 11.02 10.04 0.39 9.74 0 ≤ π ≤ 0.15 Tight 4.00 17.94 7.74 9.95

• 0.50

7.08 No info on p Beta Flat 3.49 2.88 10.76 10.54 5.44 10.54 1000 uniform 3.39

• 0.05

5.69 10.29 2.91 9.41 0.07 ≤ π ≤ 0.17 Tight 3.16

• 6.85

6.89 10.76 7.65 10.92 0 < p ≤ 0.0005 Beta Flat 3.32

• 2.12

3.73 10.30 3.01 7.05 uniform 5.00 47.46 11.16 9.99

• 0.09

6.61 0 ≤ π ≤ 0.15 Tight 4.98 46.78 6.46 10.06 0.61 5.94 No info on p Beta Flat 4.89 44.36 16.05 10.22 2.20 6.56 5000 uniform 3.47 2.39 4.11 10.10 0.97 6.94 0.07 ≤ π ≤ 0.17 Tight 3.43 1.21 6.00 10.28 2.77 6.80 0 < p ≤ 0.0005 Beta Flat 3.46 2.05 2.47 10.22 2.17 6.41 32

Effect of current sampling interval (mc)

• In general, we observe that as mc increases, the precision (inverse of the

sampling error) of the estimate of p decreases, whereas that of π increases. An intuitive explanation can be provided for such behavior. If mc = 1, i.e., every item is sampled, then one would obtain a very precise estimate of p, but a poor estimate of π, since the very first defect will be detected. The reverse occurs if mc is very large.

• Note that in Table 31 we only include the MCMC estimates, since the MAP

estimates are not recommended on the basis of the results in Tables 3 and 4.

33

Effect of sample size (N)

• Although from a practical point of view, the issue of sample size is usually

important in most estimation problems, it may not be very important here, since the estimation procedure does not require generation of fresh data; rather needs only historical data. The quality department of any organization will have plenty of past inspection records which can be used for estimation.

• Clearly, it is difficult to provide a theoretical solution to the problem of

determination of optimal sample size in the present context. To obtain an idea about the optimal sample size in the current simulation example, we study the effect of sample size on the mean squared error (MSE) of the estimates.

• Now, we vary the sample size N (number of cycles) from 10 to 300 in steps
• f 10, and compute the MSE as the sum of squared absolute bias and
• variance. Figure 35 shows the plot of estimated MSE( ˆ

p) versus N, corresponding to situation 1 with a uniform prior.

34

Effect of sample size (N)

• A sample size of about 150 should be good enough, since the reduction in MSE after

150 is not very significant. MSE(ˆ π) (not shown here) on the other hand, stabilizes much faster. The results are similar for situation 2 and the beta priors.

50 100 150 200 250 300 8.0 e−09 1.0 e−08 1.2 e−08 1.4 e−08 1.6 e−08 N(sample size) MSE of estimated p

35

Estimation of parameters in Case III

• Case III, discussed by Nandi and Sreehari (1997), deals with a scenario

where there are two types of assignable causes, termed minor and major, and their appearances follow geometric patterns with parameters p1 and p2.

• Occurrence of a major assignable cause leads to a situation like Case I,

where all subsequent items produced are defective.

• A minor assignable cause leads to a situation like Case II, i.e., the process

starts producing 100π% defective products following the occurrence of such a cause.

• Although Nandi and Sreehari (1997) derived expressions for the expected

loss, they completely ignored the estimation of p1, p2 and π.

36

The Bayesian estimation of p1, p2 and π

Note that, for this case we can use exactly the same notations as in Case II if we define p = p1 + p2 − p1p2, i.e., q = q1q2, where qi = 1− pi for i = 1,2. To develop the Bayesian algorithm we need the following result: PROPOSITION 3 : The probability mass function of S is given by P(S = s) = (1−qm)

• α(1−π)(πqm

2 +1−qm 2 )∆

{(1−π)qm

2 }r−1 −qm(r−1)

(1−π)qm

2 −qm

• +

qm(r−1) 1−∆α(1−π)

• ,

s = 1,2,...,∞. where r = s−[l/m], p = p1 + p2 − p1p2, and ∆ = q2. 1−q1 1−qm

1

.qm

2 −qm 1

q2 −q1 . We assume that pi ∼ Beta(αi,βi),i = 1,2 and π ∼ Beta(α3,β3). Then Proposition 3 leads to the following corollary.

37

The Bayesian estimation of p1, p2 and π

logg(p1, p2,π|s1,...,sN) ∝ N log(1−qm) +

N

∑

i=1

log

• α(1−π)(πqm

2 +1−qm 2 )∆{(1−π)qm 2 }ri−1 −qm(ri−1)

(1−π)qm

2 −qm

+qm(ri−1){1−∆α(1−π)}

• +

2

∑

j=1

(α j −1)log p j +(α3 −1)logπ+

2

∑

j=1

(β j −1)logq j +(β3 −1)log(1−π)

• Assuming that we have some lower and upper bounds for each of the three

parameters, the hyperparameters αi,βi, i = 1,2,3 can be obtained in the same way as discussed before.

• If the engineers are more or less certain about the limits and are unable to say

anything more about the prior distributions, uniform priors could be a possible choice again.

• Considering the complication involved in finding the posterior modes, we only use

MCMC methods to simulate the posterior density of each parameter.

38

Simulation results

We consider the same numerical example as the one used by Nandi and Sreehari (1997) where π = 0.10, p1 = 0.001 and p2 = 0.002. As in Case II, we generate 200 cycles in each simulation of the process. The following two levels of prior information are considered:

• 1. “Strong” (Reasonably accurate information) :

0.001 ≤ p1 ≤ 0.003,0 < p2 ≤ 0.002,0.07 ≤ π ≤ 0.13

• 2. “Weak” (Moderately accurate information) :

0 < p1 ≤ 0.005,0 < p2 ≤ 0.003,0 < π ≤ 0.25 In order to study the sensitivity of the method with respect to the choice of the prior distributions, we consider the following three priors: (i) Beta priors tightly distributed in the stated intervals with γp1 = γp2 = γπ = 0.05. (ii) Flatter Beta priors with γp1 = γp2 = γπ = 0.25. (iii) Uniform priors in the stated intervals.

39

Case III: Summary of Simulation Results

100 simulations were carried out for each of the 6 cases. Each estimate is the median of its simulated posterior distribution (10,000 MCMC iterations with burn-in of 1000).

Level of Prior Distribution π p1 p2 information ×10−2 ×10−3 ×10−3 Mean 10.11 2.08 0.88 High belief Beta Median 10.10 2.07 0.88 STRONG α1 = 22,α2 = 4,α3 = 40, Mode 9.99 2.06 0.88 β1 = 9980,β2 = 4000,β3 = 350 Rel Bias (%) 1.1 4.0

• 12.0

sd 0.43 0.22 0.14 Mean 9.69 2.06 0.93 0.001 ≤ p1 ≤ 0.003, Low belief Beta Median 9.59 2.04 0.94 0 ≤ p2 ≤ 0.002, α1 = 5,α2 = .55,α3 = 12, Mode 9.27 2.03 0.94 0.07 ≤ π ≤ 0.13. β1 = 2500,β2 = 400,β3 = 110 Rel Bias (%)

• 3.1

3.0

• 7.0

sd 1.14 0.38 0.24 Mean 9.82 2.00 0.95 Median 9.75 2.00 0.97 Uniform Mode 9.60 2.06 1.04 Rel Bias (%)

• 1.8

0.0

• 5.0

sd 0.74 0.37 0.22 40

Case III: Summary of Simulation Results (Contd.)

Level of Prior Distribution π p1 p2 information ×10−2 ×10−3 ×10−3 Mean 10.00 2.17 0.94 High belief Beta Median 9.71 2.10 0.94 WEAK α1 = 22,α2 = 4,α3 = 40, Mode 8.86 2.09 0.95 β1 = 9980,β2 = 4000,β3 = 450 Rel Bias (%) 0.0 8.5

• 6.0

sd 2.88 0.55 0.22 Mean 10.05 2.43 0.89 0 ≤ p1 ≤ 0.005, Low belief Beta Median 8.87 2.26 0.91 0 ≤ p2 ≤ 0.003, α1 = 5,α2 = .55,α3 = 12, Mode 6.88 2.15 0.99 0 ≤ π ≤ 0.25. β1 = 2500,β2 = 400,β3 = 110 Rel Bias (%) 0.5 21.5

• 11.0

sd 4.83 0.77 0.31 Mean 9.63 2.45 0.94 Median 8.68 2.27 0.95 Uniform Mode 7.32 2.15 1.00 Rel Bias (%)

• 3.7

22.5

• 6.0

sd 4.32 0.75 0.29 41

Comments on Simulation Results

• 1. The method works satisfactorily, even when the prior information is not

quite accurate.

• 2. However, as expected, the accuracy of prior information increases the

efficiency of the parameter estimates. This is supported by the fact that with “strong” prior information the variances of the estimators are much less that with “weak” information.

• 3. The method does not seem to be very much sensitive to the choice of the

nature of prior distribution, as also seen in Case II. If the prior information

• n π is of the form π0 ±δ, then it might be easy to elicit a Beta prior with

mean close to π0. On the other hand, if it is simply of the form of an interval πL,πU, it would be more pragmatic to consider a uniform prior.

42

Summary and Conclusions

• Noting that the estimation problem is trivial for Case I, we highlight the

problems associated with the estimation of the process parameters in Case II (p and π) and Case III (p1, p2 and π).

• We propose two different estimation procedures to resolve the

aforementioned problems. One is based on the Bayesian approach and the

• ther based on the EM algorithm.
• We propose some concrete guidelines for eliciting a prior distribution from

the available information.

• One interesting area of future research in this area is to develop a generic

framework with k types of assignable causes that would have Case I, Case II and Case III as special cases. This is encountered in several industrial situations.

43

Future Research

• This paper considers situations where neither type I nor type II errors are

likely to occur. The control problem discussed here is therefore different from the classical control charting problem.

• The inspections need not be error-free always. We are working on a

situation where the inspection may be imperfect. Borges et al. (2001) has investigated such a case using a similar model.

44

Thank you

45