You are on page 1of 16

Journal of Optimization in Industrial Engineering 19 (2016) 9-24

Cost Analysis of Acceptance Sampling Models Using Dynamic
Programming and Bayesian Inference Considering Inspection Errors
Mohammad Saber Fallah Nezhada,*, Abolghasem Yousefi Babadib
a

Associate Professor, Industrial Engineering Department, University of Yazd, PO Box 89195-741, Yazd ,Iran
b
PhD student of Industrial Engineering, Tehran University, Tehran, Iran
Received 12 January, 2015; Revised 07 September, 2015; Accepted 01 October, 2015

Abstract

Acceptance Sampling models have been widely applied in companies for the inspection and testing of the raw materials as
well as the final products. A number of lots of the items are produced in a day in the industries so it may be impossible to
inspect/test each item in a lot. The acceptance sampling models only provide the guarantee for the producer and consumer
confirming that the items in the lots are according to the required specifications so that they can make appropriate decision
based on the results obtained by testing the samples. Acceptance sampling plans are practical tools for quality control
applications which consider quality contracting on product orders between the vendor and the buyer. Acceptance decision is
based on sample information. In this research, dynamic programming and Bayesian inference is applied to decide among
decisions of accepting, rejecting, tumbling the lot or continuing to the next decision making stage and more sampling. We
employed cost objective functions to determine the optimal policy. First, we used the Bayesian modelling concept to
determine the probability distribution of the nonconforming proportion of the lot and then dynamic programming was utilized
to determine the optimal decision. Two dynamic programming models have been developed. The first one is for the perfect
inspection system and the second one is for imperfect inspection. At the end, a case study is analysed to demonstrate the
application the proposed methodology and sensitivity analyses are performed.
Keywords: Acceptance Sampling, Bayesian Inference, Dynamic Programming, Inspection Errors, Quality Cost.

1. Introduction
stochastic dynamic programming has the same approach
of a deterministic one, but only the state transition
equation differs. The acceptance sampling problem may
be modelled as a dynamic programming problem when
different sampling stages are available. States of the lot
may be defined by the results of the inspection. The lot
state is defined as the expected value of nonconforming
proportion. The probability distribution function of
nonconforming proportion is obtained using the Bayesian
inference. Therefore, the lot state is assumed to be known
at each stage and the probability density function of
nonconforming proportion is determined at the beginning
of each stage after sampling the new data.
The state of the lot at the beginning of the next
stage depends only on our current decision (tumbling or
more sampling). The lot can be tumbled a constant cost.
Tumbling will bring the lot to some better state in the next
stage and it decreases lot state (nonconforming
proportion) with a constant factor. There is also a statedependent cost of decisions about accepting and rejecting

Acceptance Sampling models have been widely applied in
companies for the inspection and testing of the raw
materials as well as the final products. A number of lots
of the items are produced in a day in the industries, so it
may be impossible to inspect/test each item in a lot. The
acceptance sampling models only provide the guarantee
for the producer and the consumer confirming that the
items in the lots are according to the required
specifications so that they can make appropriate decisions
based on the results obtained by testing the samples.
In this paper, an optimization model is developed
for acceptance sampling plan. The proposed approach is
based on dynamic programming and Bayesian inference.
In deterministic dynamic programming, given a state and
a decision, both the immediate payoff and next state are
known. If we know either of these only as a probability
function, then it is modelled as stochastic dynamic
programming. The method of obtaining stages, states,
decisions, and recursive formula does not differ. A
* Corresponding author Email address: Fallahnezhad@yazd.ac.ir

9

The credit of the producer was defined as the total number of items accepted since the last rejection. Moskowitz and Tang (1992) proposed acceptance sampling plans based on the Bayesian inference. or reject the lot if the observations are ‘‘far’’ from the specification limit(s). Fallahnezhad & Niaki. Since an inspection process may be imperfect thus inspection process leads to a biased estimation of nonconforming proportion hence the first correct value of nonconforming proportion should be evaluated using conditional probability. in comparison with single sampling plan or double sampling plan. They assumed that the inspection is not perfect and type I and type II errors occur in the inspection process. 2015. accepting or rejecting the lot. Fallahnezhad and Aslam. 2013. The problem of acceptance sampling plan in the presence of inspection errors is modelled using decision tree approach. Fallahnezhad (2014) and Fallahnezhad and Niaki (2011) proposed an optimal policy for machine replacement problem in a finite horizon model based on the quality of items produced. In order to determine the optimal policy. Tagaras (1998) proposed a dynamic model for the joint process control and machine maintenance problem of a Markovian deteriorating machine. and provided numerical solution to the sequential sampling plan using Bayesian approach and dynamic programming. Different decisions are employed in objective function. Fallahnezhad et al. Lindley and Barnett (1965) assumed that the probability for an item to be nonconforming has a beta prior distribution. 2014. They used from Bayesian inference to update the probability distribution function of nonconforming proportion. Aslam & Fallahnezhad. reflecting the final decision about the lot. the lot at the end of the horizon. and a suitable cost model is employed depicting the cost of sampling. (2013) proposed a new approach on selecting the different actions for a lot based on Bayesian modeling. Then they analysed the cost at terminal nodes and optimal decisions are determined using a backward recursive approach. the inspection cost in combination with the cost of continuing the production process and cost of rejecting was minimized. we propose an adaptive optimal policy for lot sentencing problem. Kuo (2006) developed an optimal adaptive control policy for joint machine maintenance and product quality control using dynamic programming. Fallahnezhad & Nasab. 2005. 2012. In this paper./ Cost Analysis of Acceptance Sampling. 2012. Fallahnezhad et al. Fallahnezhad et. This policy is derived based upon a dynamic programming and the Bayesian approach while inspection may be imperfect and first and second type errors existed in the inspection process. 2013.. Niaki and Fallahnezhad (2009) used Bayesian inferences concept and stochastic dynamic programming to design an acceptance sampling plan. Also Fallahnezhad et al. They used a stochastic dynamic programming model to design a control threshold policy for the machine replacement problem. we shall accept the batch if all observations up to date are ‘‘close’’ to the specification limit(s). an optimal framework is developed for the decisionmaking process at hand. Iravani and Duenyas (2002) considered integrated decisions of maintenance and production in a production system with a deteriorating machine. (Li & Chang. They used a stochastic dynamic programming model to minimize the ratio of the system cost to the system correct choice probability. Fallahnezhad and Yousefi (2014) developed a new acceptance sampling plan to decide about the received lot based on cost objective function in the presence of inspection errors. (2012). Klassen (2001) introduced an acceptance sampling system based on a new measure named ‘credit’. They applied Bayesian modeling to determine the probability distribution of the nonconforming proportion of the lot and expected cost of different actions. Using a Bayesian procedure. The rules for a sequential sampling plan are as follows: at each time after inspection of one lot. Even though several dynamic programming models had been proposed for the problem of acceptance sampling but as the best of author's knowledge. Case studies and sensitivity analysis are performed to elaborate the application of proposed model.. the dynamic modeling of this problem based on a continuous state variable and proposed definition for sate variable is not addressed before. or continue sampling one more items from the lot 10 . Fallahnezhad & Nasab.. Also solution method of dynamic model is a new heuristics method. (2012) proposed a decision tree approach for acceptance sampling model. For attribute sampling plan. They proposed a new acceptance sampling design to accept or reject a lot based on Bayesian modelling in order to update the probability distribution function of the nonconforming proportion. a new dynamic programming model is developed for acceptance sampling problem. Then the developed model is generalized for the imperfect inspection system. They consider infinite horizon and assumed that the production and repair times follow exponential distribution. At the end. a sequential sampling plan is more attractive in terms of statistical efficiency.al. On the other hand. The objective is to minimize the total cost over the horizon of decision making. There are various models for designing an economically optimal sampling system. In this research.. Niaki and Fallahnezhad (2007) utilized Bayesian inference and stochastic dynamic programming to design a decision-making framework in production environment. 2011). a prior distribution is specified. They presented a procedure based on dynamic programming for designing optimal acceptance sampling plans for item-by-item inspection. Ivy and Nembhard (2005) proposed statistical quality control (SQC) and partially observable Markov decision processes (POMDP) for maintenance decision making of deteriorating systems. then corresponding costs are evaluated based on the correct value of nonconforming proportion.Mohammad Saber Fallah Nezhad et al. Wortham and Wilson (1971) proposed an optimal sequential sampling plans using backward recursive inference. Fallahnezhad MS.

 State variable: The expected value of nonconforming proportion. Vn ( p) : The cost associated with p when there are n remaining stages to make the decision. On the other hands dynamic programing models are solved based on the recessive approach. before doing so.  We can select the optimal policy among decision of accepting. ii) The proportion of nonconforming items is stable and constant or its probability distribution is known. The Model the lot is accepted.  Dynamic programming is used to find the optimal policy.5) p   0. In this paper we impose mandatory fixed sample sizes in each sampling stage.5 . A dynamic programming algorithm will examine all possible methods to solve the problem and will select the optimal solution. I : The cost of inspecting one item during decision of inspecting all items in lot. 1983). Stochastic dynamic programming is a technique to model a sequential decision making process in a stochastic environment (Ross.5) (   0.5 (1  p )   0. C a : The cost of accepting one nonconforming item when 2. The state variable and stage variable of dynamic programming model is as follows. therefore. Following assumptions are made to design the proposed sampling plan.  The inspections are perfect. However. rejecting.  : The discount factor in stochastic dynamic programming approach. (2001)). We let the dynamic programming mechanism dictate the optimal policy based on the current state of the system. al. Dynamic programming technique can be employed to design an optimal sequential acceptance sampling plan when the following conditions exist: i) The cost of accepting a nonconforming item and cost of rejecting a conforming item can be reasonably assessed. C s : The cost of one nonconforming item which is detected during inspection.Journal of Optimization in Industrial Engineering 19 (2016) 9-24 otherwise. A recessive approach was used to design these models.  A tumbling operation can be performed on lot. T : The cost of tumbling process. It is assumed that there are n max decision making stages.making process. Derivations We may model an acceptance sampling process as an optimal stopping problem in which in each stage of the decision-making process.  The objective function minimizes the summation of quality costs. dynamic programming enables us to go through all possible solutions to select the best one.5 and 0.5 (1) Where. first we need to have some notations and definitions. Notations and Definitions We will use the following notations and definitions in the rest of the section: p : The proportion of nonconforming items. we take a sample from a lot and based on the information obtained from the sample we want to decide whether to accept or to reject the lot or continuing to take more samples. tumbling the lot or continuing to the next stage and taking more samples. therefore we use from the dynamic programming to develop a sequential sampling models.1. 2. The probability distribution of nonconforming proportion is obtained by sampling. we are selecting between decisions of continuing (tumbling the lot or continuing to the next decision making stage and more sampling) or stopping (accepting or rejecting the lot) thus it is a type of optimal stopping problem that can be generalized in order to consider all decisions.  is the number of nonconforming items and  is the number of conforming items in the past stages of the decision. we take a Beta prior distribution with parameters 0. We mentioned that the probability distribution of the nonconforming proportion ( p ) could be modelled by the 2. N : The total number of items a lot. Referring to Jeffrey’s prior (Nair et.  Bayesian inference is used to update the proportion of nonconforming items. By use of the Bayesian inference.  Stage variable: The stage of sampling. The most powerful method of acceptance sampling plans is sequential acceptance sampling model. In acceptance sampling plans. The tumbling operation can be expected to eliminate 1  percentage of the nonconforming items. m : The sample size in each stage of decision making. we can easily show that the posterior probability density function of p is f (p)  (    1) (  0. for the nonconforming proportion p.  : The coefficient of decreasing detective proportion after improving lot quality (Tumbling). 11 .2.

Thus. then a 100% inspection plan is done on the lot and all items are inspected thus the total cost of NI  C s NE  p   n  0 then.   0. when no decision making stage is available  n (6) based   NI  C Np  f  p dp  The discount factor  is used to evaluate the cost of the next stage in the current stage. The approximate values of parameters  '. then the proportion of nonconforming items in the next stage remains p and if tumbling decision is made in one stage then the proportion of nonconforming items in the next stage decreases from p to  p where  1.C a NE ( p ) (3) It is seen that the value of V n ( E ( p )) can be easily computed from the values of V n 1 (.) .5 . '   0.C a Np  Thus following is concluded approximation method in equation (5). NI  C S NE ( p )  Ca NE ( p )  Accept  (4) n 0   NI  C S NE ( p )  Ca NE ( p ) Reject  incurred and decision making process ends. Bayesian inference as a Beta distribution with parameters   0.  ' ..  T + V n 1  E  p1   p    we should select between decisions of equalities. We characterize properties of the value function at the last stage using a method for approximating the function V 0 ( E (.Mohammad Saber Fallah Nezhad et al. then V 0 (E ( p ))  Min NI  C s NE ( p ). Assume that p1 follows a Beta distribution with  (2) parameters  './ Cost Analysis of Acceptance Sampling. It is obvious that on stage n  1 we may choose one of the possible decisions again.  f ( p1 )  Beta ( ' . therefore V 0 ( E ( p )) is determined based on the minimization of two terms. is accepting or rejecting the lot. when no decision making stage is available NI  C s NE  p  shows the cost when we reject the lot. we have only two decisions (accepting and rejecting the lot). It is assumed that if lot continues to the next stage. thus we may easily compute all cost values recursively based on the values of V 0 (. It is obvious that new sample data is gathered in this case and then the probability distribution function of the state variable is updated using Bayesian method and dynamic model would be solved again with new sample data. This factor can be addressed as the rate of decreasing the cost at the next stages which may be resulted from having more information on the state variable in the next stages. then we should select between decisions of accepting or rejecting the lot that can be easily performed by comparing their corresponding cost. This probability distribution function is needed in equation (7) for the cases that tumbling action has been selected.) . and mI   V n 1  E ( p )   shows the cost when we continue to the next decision making stage therefore sampling continues.5 .C a NE ( p ) . thus following is obtained. It is assumed that when we reject the lot.. Hence the cost associated with p when there are n remaining stages to make the decision is:     (5) Since V 0 ( p ) can be obtained for all possible value of p using following equation.  ' ) 12 (8) . C a N E  p  represents the cost when we accept the lot. T   V n 1  E ( p1   p )   NI  C s NE ( p ). V 0  E  p    E V 0  p    s p on (7) I C a C s  C Np  f  p dp a p I C a C s Also the probability distribution function of the random variable p1 is needed to evaluate function V n 1  E  p1   p   . Tumbling operation can be expected to eliminate 1   percentage of the nonconforming items therefore the proportion of nonconforming items in the tumbled lot reduces to p1   p thus Since p is a stochastic variable.5    E  p     E  p1     '  '     1   '  '         0  . thus we have applied approximation method in equation (5) in order to consider the probability distribution of p in the computations.)) as follows: V 0  E  p    E V 0  p   represent the cost when we tumble the lot.  ' can be determined from the following Also on stage n  0 . V 0 ( p )  Min NI  C s Np . Thus. This probability distribution is determined using a heuristic approach. If we define n to be the index of the decision-making stage and p to be the state variable. Also a Tumbling operation can be performed on lot.mI  V n (E ( p ))  Min  V n 1  E  p    .

V 2  E ( p )  can be calculated using recursive equation . the cost of one nonconforming juice which is detected during inspection is C s  2 .  can be obtained using equation (2) recursively thus it is concluded that the value of V n ( E ( p )) can be determined based on the  NI  C s NE  p  . the cost of tumbling process is T  150 . Steps of dynamic programming have been shown in Fig (1). 10 juices are inspected where the number of nonconforming juices is   2 and the number of conforming juices is   8 in the first sample and the . Assume that stages are available for decision making process and at the start of process.  NI  C s NE  p  . According to the presented approach. 1. first a sample of items is inspected. Also three decision making stages are available for deciding about the lot. on the other hand we have : Also the value of V 0  E (.C a NE  p   V 1  E ( p )   Min mI   V 11  E  p   .     follow:  NI  C s NE  p  .     when two decision making stages is available. The sample size in each stage of decision making is m  10 .  T   V 2 1  E  p1   p   Now we can recursively determine V 1  E ( p )  as 3.C a NE  p   V 2  E ( p )   Min mI   V 2 1  E  p   . tumbling process is According to dynamic programming approach when three decision making stages is available ( n  3) .9 .  T   V 11  E  p1   p     2 Also we need to obtain V 1 E p 2   p1   p calculating the item V 2 E  p 1   1 2 need to determineV 0 E p 2   p1   p item is I  1 . a case study is given to illustrate the application of the proposed methodology.  .  ' can be determined based on the values of   95% and the discount factor in stochastic dynamic programming approach is   0.       and the   which are . The amount of vitamin C in juice is inspected through experimenters. C a NE  p   V 3  E  p    Min  mI   V 3 1  E  p   .)  can be determined by equation (7). Diagram of Dynamic Programming’s steps In the next section.)  by continuing this recursive method. 2 3 function V 0 E p 3   p 2   p1   p 13   for   p   and then we cost of accepting one nonconforming juice is C a  4 . Therefore the values of parameters  '. Now we can evaluate the value of function V n ( E ( p )) based on the value of V n 1  E  p   and V n 1  E  p1   p   where V n 1  E . Case Study Assume a juice production industry has produced a lot of N  100 items. the coefficient of decreasing nonconforming proportion in obtained by equation (6).Journal of Optimization in Industrial Engineering 19 (2016) 9-24 The second equality is defined based on constant sample size in each stage. Fig. we have : parameters  .  T  V 3 1  E  p1   p   values of V 0  E (. the cost of inspecting one   .

first we must calculate the V 0  E ( p )  . In general.) E p 0.483 227. then V 3  E ( p ) as reported in Tables (1)…(5).684 NI  Cs NE  p1  Table 2 Results of calculating V n  E ( p )  n=0 V 0  E ( p1 )  0.2272727 E  p1   p    E  p  0. V 2  E ( p )  and at the end Table 1 Results of calculating E(.3636 Ca NE  p  90.9091 T   V 0  E  p3  252. 14 ..7273 90.455 V 2  E ( p1 )  86.182 0.841 NI  Cs NE  p  145.727 V 3  E ( p ) 90..115 137.19486 Ca NE  p1  86.3636 Ca NE  p  90. sensitivity analysis is performed on different parameters.9091 T   V 1  E  p2   223.233 mI   V 2  E  p   91.799 T   V 0  E  p2  257.8182 V 1  E ( p2 ) 82. Sensitivity Analysis A sensitivity analysis is performed on the parameters of the problem that results are in the Table (6).535 NI  Cs NE  p  145. the optimal policy is to accept the lot when three decision making stages is available.3636 mI   V 1  E  p1   87.3636 mI   V 0  E  p1   123.535 n=0 NI  Cs NE  p  145.20511 E p 2   p1   2 p   2 E  p  3 Table 3 Results of calculating V n  E ( p1 )  263.Mohammad Saber Fallah Nezhad et al./ Cost Analysis of Acceptance Sampling.455 V 1  E ( p1 )  86.9091 Ca NE  p2  82.0455 T   V 2  E  p1   p   227.9091 NI  Cs NE  p1  143.0455 mI   V 0  E  p 2   117. V 0  E ( p3 ) 113. In the next section.99 Ca NE  p1  86.9091 n=1 Table 5 Results of calculating V n  E ( p 3 )  n=0 Considering Tables (1-5).2159091  E p  n=0 n=1 143.592 4.115 n=2 Table 4 Results of calculating V n  E ( p 2 )  E V0 ( p2 )  119.8182 T   V 1  E  p1   p   V 2  E ( p )    p 2   2 p1   3 p   3 E  p  n=1 V 0  E (p ) T  V  E p 0 1   p  V1  E ( p ) n=2 n=3 125.023 90.727 NI  Cs NE  p2  141. thenV 1  E ( p )  .455 Ca NE  p  90.9091 mI   V 1  E  p   91.182 mI   V 0  E ( p )  128.

 .1. 2.150. 2.13.C s . 4.95.722773 7. 2. 4. I . 2.T .1. Decreases  (100.836 3. 2.95.544 4.10.150.8) Accept 90. 4. 2.5. confirms the reasonable performance of dynamic programming model in the encountering with the variation of cost parameters.1.9. Increases N (3422. 2.95. 0.10. 2.95. 0.150.5.150.1. m . 2.8) Tumble the lot 90. 2. 0. when the nonconforming proportion is more than 0.150. For example.8) Continue to the next decision making stage Continue to the next decision making stage 6. 0.150. 0.95.9. Decreases C s (100.9.9.10. it is observed that the optimal policy is a type of control threshold policy. 0. When the nonconforming proportion is more than 0.  By comparing case one and case six. it is seen that when the cost of tumbling process T  is less than 13.95. 2.10.150. 4.53. 4. Case study (100.8) Tumbling 46. then the optimal decision in the proposed method is to continue to the next decision making stage. Sensitivity analysis of “  : the number of nonconforming items. 0.5.10. 4. The results of changing  and  are denoted in Table 8. 4. 0. 0.  : the number of conforming items”:  The result of the proposed model in all cases is reasonable.95. two and three.54983 9. 2. 0.9091 2.8) 83. C a .10.02635 166. it is seen Sensitivity analysis of variations of “ C a and C s ” N. 2.31. cost of proposed model increases by increasing cost parameters of the model and it decreases by decreasing different costs of the model and the corresponding optimal decision are made in accordance with the variations. then optimal policy is to continue to the next decision making stage therefore sampling continues.2727 88.O of Cases optimal policy is Ca  C s  Accept the lot less than 3422 units. then optimal policy is to accept the lot.9.67 then the optimal decision is to reject.1. 2. The results of Table 7. 0. and when the total number of items in Ca  C s  Tumbling Ca  C s  Continue to the next decision making stage Ca  C s  Accept the lot that when the total number of items in a lot a lot N  N  is more than 3423 units.8) Tumble the lot 2539.8) Continue to the next decision making stage 2538. 0.150. 0.1.2. 0. 0. Decreases I (100.25.13636 8. 2. 2.  .9. then the optimal decision in the proposed method is to tumble the lot. Decreases T (100. Thus. Increases C a (100.95. 4.1.9.48381 Table 7 A sensitivity analysis is performed on the parameters of the problem that results have been summarized as following:  By comparing cases one.1.31. 0. Decreases  Continue to the next decision making stage (100. 2. Sensitivity analysis of “ C a with C s ”: Results for Simultaneous variations of cost of C a  and the cost of one detected nonconforming item C s  are denoted in accepting one nonconforming item 15 .  ) Optimal policy V 3  E ( p ) 1.10.10.8) Continue to the next decision making stage 34. then the optimal decision in the proposed method is to tumble the lot.Journal of Optimization in Industrial Engineering 19 (2016) 9-24 Table 6 Sensitivity analysis cases (N . 0.1. Increases N (3423.8) 5.10. 0.  . It is seen that when the nonconforming proportion  p    is approximately equal or less       than 0. 0. 4.9. 4. 0.

Table 8 Sensitivity analysis of “  and  ” m=9   p optimal policy m=10   p optimal policy m=11   p optimal policy m=12   p optimal policy m=13   p optimal policy m=14   p optimal policy m=15   p optimal policy 1 8 0.11 Accept 1 a 2 9 b 8 0.10 p 0..29 Accept optimal policy Accept 3 a 4 12 b 11 0. 0.08 Accept 2 11 0. 0.00.40 Continue 4 7 0..27 Accept optimal policy Continue 2 10 0.21   p optimal policy   p optimal policy   p 3 9 0.67 Reject .31 Accept 4 10 0./ Cost Analysis of Acceptance Sampling.67.38 Continue 5 9 0.67)  continue to the next decision making stage  p  [0.33 Continue   p optimal policy 10 5 0.36 Continue 5 10 0.Mohammad Saber Fallah Nezhad et al.  p  [0.2 p 0.36 Continue 4 8 0.1.25 Accept 4 9 0.3 Accept optimal policy Continue 2 a 3 9 b 8 0.33 Continue 5 8 0.20 Accept optimal policy 1 10 0.27 Accept optimal policy Accept Consequently following decision making strategy is resulted.09 Accept 1 11 0.23 Accept 3 11 0.15 Accept 2 12 0.00)  reject the lot  16   p optimal policy   p optimal policy   p optimal policy   p optimal policy   p optimal policy   p optimal policy 4 6 0.31.13 Accept   p optimal policy   p optimal policy   p optimal policy   p optimal policy   p optimal policy   p 3 7 0.18 p 0.14 Accept 2 13 0.31]  accept the lot   p  (0.17 Accept 3 10 0.

First. we take a sample from a lot in each stage of the decision-making process and the objective is to decide whether to accept. I 2 (Probability of accepting one nonconforming item). We may model an acceptance sampling process as an optimal stopping problem.5 . then the cost functions of different decisions can be obtained as follows:  item is categorized   item is   item is   pr       as nonconforming   conforming   conforming  p  pr   item is categorized   item is  item is   pr   pr       as nonconforming   nonconforming  nonconf orming  17 (10) . 1  I 2 (Probability of rejecting one nonconforming item). or to continue sampling based on the information obtained from the sample. But ( p ) denotes * * * a2  NI  NI 1C r (1  p )  NI 2C a p  C s N (1  I 2 ) p  a2  NI  NI 1C r  ( NI 2C a  NI 1C r  C s N (1  I 2 )) p * 3) for tumbling the lot * a3  T  V n 1 ( p T ) 4) for continuing to the next decision making stage and more sampling * a4  ms  V n 1 ( p ) (9) It is assumed that when the lot is rejected then all items are inspected and three types of cost are incurred. p : The apparent nonconforming proportion (its value is obtained by imperfect inspection). I1 : The probability of first type error in inspecting one Np* is the number of nonconforming items in the * lot. p * : True proportion of nonconforming. thus its value is different from the true nonconforming proportion. p T : The apparent proportion of nonconforming items in the tumbled lot. If we define n to be the index of the decision- is determined using conditional probability as follows. nonconforming item during inspection multiplied by probability of detecting a nonconforming item. * making stage and E ( p ) to be the state variable. Cr : The cost of rejecting one conforming item. C s N (1  I 2 ) p * : The cost of one detected : The cost of accepting one nonconforming item multiplied by second type error probability. tumbled lot. the second type of error. True value of nonconforming proportion ( p ) item. to reject and inspect all items. 1. the nonconforming proportion obtained by imperfect inspection. These costs are as follows. I1 (Probability of rejecting one conforming item). There are two types of inspection errors: 1. 3.   0.Journal of Optimization in Industrial Engineering 19 (2016) 9-24 1) for accepting the lot a1  C a N 5. NI 2C a p * 2. We mentioned that the probability distribution of the nonconforming proportion ( p ) could be modelled by the Bayesian inference as a Beta distribution with parameters   0. I2 : The probability of second type error in inspecting one item. The first type of error 2. Following notations are used in the rest of this paper. to tumble the lot. S : The cost of inspecting one item during sampling. p *T : True proportion of nonconforming items in the NI 1C r (1  p * ) : The cost of rejecting one conforming item multiplied by first type error probability.5 . The Presence of Inspection Error 2) for rejecting the lot and inspecting all items in the lot In this section. a dynamic programming model is developed for acceptance sampling problem in the presence of inspection errors.

   1   0. Also a discount factor  is needed to evaluate the cost of the next stage in the current stage (according to the approach of stochastic * dynamic programming). Thus following is obtained. then we can select one of the possible decisions again. Since.. and if tumbling decision is made in 18 .5 . p  I1 (1  p* )  (1  I2 ) p*  p*  Since E ( p )  p  I1 1  I1  I 2 (12)   0. we have only two decisions of * accepting and rejecting the lot therefore V 0 ( E ( p )) is determined based on the minimization of two terms./ Cost Analysis of Acceptance Sampling. * stage remains p .  item is categorized   item is   pr       I1  as nonconforming   conforming   pr item is conforming  1  p* (11)    item is categorized   item is pr     1  I 2  as nonconforming   nonconforming   pr item is nonconforming  p* Thus.5 I E ( p )  I 1    1 1 * E (p )   1 I1  I 2 1 I1  I 2 Tumbling operation eliminates 1    * items in the tumbled lot reduces to p T ( pT* ) (13) percent of the nonconforming items therefore the proportion of nonconforming   p * .5 I E (p )  I1    1 1 * * * * * p T   p  E ( p T )  E ( p )   E ( p )    1 I1  I 2 1 I1  I 2 (14) It is obvious that when n  1 stages are available. So the true value of nonconforming proportion after the tumbling process is determined as follows. Hence the cost associated with E ( p ) when there are n remaining stages to make the decision is: a1  C a NE ( p * )    * a2  NI  NI 1C r  (NI 2C a  NI 1C r  C s N (1  I 2 ))E ( p )  * V n (E ( p ))  Min   * a3  T  V n 1 (E ( p T ))  a  ms  V (E ( p * ))   4 n 1  * a1  Ca NE ( p ) . When no decision making stage is available (15) (16)  n  0 . a2  NI  NI 1C r   * V 0 (E ( p ))  Min   * (NI 2C a  NI 1C r  C s N (1  I 2 ))E ( p ) Since in stage n  0 . then the nonconforming proportion in the next then we must select between two decisions of accepting or rejecting the lot that can be easily performed by comparing their corresponding cost.   0. It is assumed that if decision making continues to next stage..Mohammad Saber Fallah Nezhad et al.

     a1  C a Np * . 1  I1  I 2 (20) thus V 0  E ( p )  is obtained as follows: V 0  E ( p * )    p  I1  N I  I C  ( I C  I C  C (1  I ))   f  s 1 r 2 a 1 r 2  1  I  I   1 2    I  I 1C r  p  C a ( I 2C a I 1C r C s (1I 2 )) 1 I1 I 2   I 2       p  I1   C a N  f  p  dp  1  I  I   1 2    I I C  p  C a ( I 2C a I1C1r r C s (1I 2 )) 1I1  I 2  I 2    When  n  0  p  dp       (21) stages are available then we must select between decisions of accepting or rejecting the lot.  "  19 "  "  " . To evaluate equation (21). first we have. distribution function of p in computations thus following approximation is applied. the probability distribution function of random variable p T is needed.Journal of Optimization in Industrial Engineering 19 (2016) 9-24 one stage then the nonconforming proportion in the next * Since function V 0 ( p * stage decreases form p to  p where 0    1 . We characterize properties of the value function using a method for approximating the function E V n -1 . thus following strategy is obtained for single stage model: N I  I 1C r  (I 2C a  I 1C r  C s (1  I 2 ))E ( p * )  C a NE ( p * )  Accept V 0  a1    n 0  * *  N I  I 1C r  (I 2C a  I 1C r  C s (1  I 2 ))E ( p )  C a NE ( p )  Reject V 0  a2       (22) follows a Beta distribution with parameters  ". This function is determined using the equation (27).  " . (23) . recursively to solve dynamic model in equation (14).  " can be  It is necessary to determineV n 1 E . Since p T is the apparent nonconforming proportion of the lot in the imperfect inspection process thus we assume that p T with parameters  ". To explain this heuristic method.  ) is obtained using equation    can be obtained using equation (18) thus E V 0 p in order to consider the probability V 0 E  p*   E V 0  p*  * * V 0 (E ( p )) can be obtained as follows:  N  I  I 1C r  (I 2C a  I 1C r  C s (1  I 2 )) p *  f  p  dp   *  I  I 1C r  p C a (I 2C a  I1C r C s (1I 2 ))       *   C a Np  f  p  dp I I C    p* C a  (I 2C a I1C1 r r C s (1I 2 ))  (19) Since.    * V 0 ( p )  Min  a2  NI  NI 1C r    * (NI 2C a  NI 1C r  C s N (1  I 2 )) p  (17) (18) hence  V 0  E ( p * )  E V 0  p *   * (19). The approximate values of parameters  ". p*  p  I1 . E  p T   Mean of Beta distribution determined using a heuristic approach.

 " can be determined using the values of parameters  ./ Cost Analysis of Acceptance Sampling. Case Study of Extended Model Thus using two equalities (24) and (25).  " .  "  "     .Mohammad Saber Fallah Nezhad et al.5  I      1 1  1 " 1   I1  1  I  I 1  I  I  "   " 1  I  I 1 2 1 2 1 2     1 therefore it is concluded that we can 6.5     "  (   ) 1    (1   ) I 1         1  Thus the values of parameters  ". assume that the cost of rejecting one conforming juice is C r  5 and the cost of inspecting one    0.1 . 20 . following is obtained.. we a1  C a NE ( p * )    a2  NI  NI 1C r   (NI 2C a  NI 1C r     * V 3  E ( p )   Min C N (1  I ))E ( p * )  2  s  a3  T  V 2  E ( p T* )     a4  ms  V 2  E ( p * )      *  V 2 E ( p ) can be obtained recursively.5   "  (   )      1  (1   ) I 1  (28)    0.  ". In the next section. "  item during sampling is S  2 and the probability of first type error in inspecting one juice is I1  0.)  recursively..      and V 0  E (. and the probability of second type error in inspecting one juice is I 2  0. we have.05 . a case study is given to illustrate the application of the extended model methodology. According to stochastic dynamic programming (26) Assuming equal number of sample size for both cases. Thus following equalities are 1 * In addition to the parameter values of the case study in the third section.5  I      1     1 I  I     '   '        determine the value of V n E ( p ) based on the value of  p  I1  E  p T*    E  p *    E   1 I1  I 2     0.5      1  I1    (24)  1 I1  I 2      Also since p T is apparent nonconforming proportion of resulted to obtain the values of   E  p  * V n E ( p ) based on the value of V n 1 E p * of the nonconforming items thus the proportion of nonconforming items in the tumbled lot reduces to 1 1 I1  I 2 I1 (27) a1  C a NE ( p * )    a2  NI  NI 1C r   ( NI 2C a  NI 1C r     * V 2  E ( p )   Min   * C s N (1  I 2 ))E ( p )  a3  T  V 1  E ( p T* )     a4  ms  V 1  E ( p * )   Following results are obtained by solving equations(27)     0.  .  p I  1 " E  p T*   E  T 1    1  I 1  I 2  1  I 1  I 2  "  " (25) 1  I1 1 I1  I 2 1 * T  the lot after tumbling obtained by imperfect inspection and p T* is the true proportion of nonconforming items after tumbling with considering inspection errors thus using equation (12) following is obtained     0. Since tumbling operation eliminates 1    Now we can evaluate the value of function percentage  * T p p * V n 1 thus approach where three stages  n  3 2 1  I 1  I 2  "  "  are remained. following is obtained: Also V  E ( p )  can be obtained recursively: * 1 . Thus when two stages are remained  n  2   have.

26 T2 7.522 * 125.169 * * 125. sensitivity analysis is performed on different parameters.738 3 T  V 0  E ( p ) * I1  increases. A sensitivity analysis is performed on the parameters of the problem that results have been explained in following: 190.133 a * NI  NI1Cr  (NI 2Ca  NI1Cr Cs N (1 I 2 ))E (p ) 119. the optimal policy is to continue to the next decision making stage and more sampling when three stages are available. In the next section. then the optimal policy in the proposed method is to reject the lot. then the optimal decision in the proposed method is to accept the lot. 2 * N  102.126 * 1  All results coincide with the type of variations. the cost of one detected nonconforming items C s  and the cost of rejecting one conforming items C r  are investigated and the results nonconforming items are denoted in Table 12. Sensitivity analysis of changing cost parameters Simultaneous variations of the cost of accepting one C a  .18822  E p T3  E  p T2  E  p T  * 3 The calculations are reported in Tables (9-10). For example.852 1 * decreases.Journal of Optimization in Industrial Engineering 19 (2016) 9-24 Table 9 a1  C a NE ( p * )    a2  NI  NI 1C r   ( NI 2C a  NI 1C r     * V 1  E ( p )   Min   * C s N (1  I 2 )) E ( p )  a3  T  V 0  E ( p T* )     a4  ms  V 0  E ( p * )   The results of E  .22727   E p   E p  E  p   E  p    E  p  E p* * T * * T2 * T 2     E p    E p  *  * * 2 * 0.738 2 * 106.852 the optimal decision in the proposed method is to tumble the lot. 0. then it’s better to accept the lot because first type error is the probability of incorrect rejection of an acceptable item. and when the total number of T * It is seen that when the total number of items in a lot becomes more than 3800 units. Considering Tables (9-10). 21 . This action is logical because when the probability of first type error in inspecting one item increases. The results show the valid and logical performance of the proposed method.839 E (p ) ms  V  E ( p )  V  E ( p ) 190.133 * NI  NI1Cr  (NI 2Ca  NI1Cr Cs N (1 I 2 ))E (p ) 119. increasing the cost of one decision leads to not selecting this decision as optimal.140 1 n=2 7.20855 0.114 T  V n=3 * 2  in inspecting one item 106. T * It is seen that when the probability of first type error 187. 103.17881 * T3 Table 10 Results of V n=0 E (p ) * n   C NE  p  * V0 E (p ) 103. then 125.1. E p 0.169 a T  V Results of sensitivity of analysis are shown in Table (11). Sensitivity analysis of extended Model  N  items in a lot 102.839  103.19812 0.133 a * NI  NI1Cr  (NI 2Ca  NI1Cr Cs N (1 I 2 ))E(p ) n=1  * T  V 0 E ( p T )  V E ( p ) C NE  p   * ms  V 0 E ( p ) 119.839 E (p ) ms  V  E ( p )  V E (p ) C NE  p  190.

2.  : the number of nonconforming items” sampling continues. 2.5. 2. then optimal p       policy is to accept the lot.5) Continue to the next decision making stage 106. 0. 0. I .  .1) (100.43)  continue to the   next decision mak ing stage  p  [0. 0..5.150. 4.1.5. then optimal policy is to reject.5. 0. 0.10.10. 0. 4.10.95. 2. 0. It is seen that when the nonconforming proportion    is equal or less than 0. 0. 0.Mohammad Saber Fallah Nezhad et al.150.0. I1 .1) (100.95.0. 2. 0.95.3544837 Table 12 Sensitivity analysis of changing two parameters N.9. 2.8.1. 2.5.  p  [0. 0. 4.9.4491979 Reject the lot 119. 2.9. 0.1) (100.18.C r .9. I2 ) Optimal policy (100. 2.05.10. 4. 2. 0. 2. 4. 2.05.  .8.5. 2. 0. 2.8.5.5.1) (100.738 Tumble the lot 2927.1) (70. 0. 0.10.1) (3880.95.05.43.95.05.10.9.10.  . 0. Sensitivity analysis of “  : the number of nonconforming items.O of cases Case study Increases N Decreases N Increases Cs Decreases T Decreases I Decreases  Increases  Decreases Cr Increases S Increases I1 Increases I2 V 3  E ( p ) (N .1. 0.00)  reject the lot The results of changing  and  are denoted in Table 13. 2.O of cases Optimal policy Ca  Cs  Accept the lot Ca  Cs  Continue to the next decision making stage Ca  Cs  Continue to the next decision making stage Ca  Cs  Accept the lot Ca Cr  Accept the lot Ca Cr  Continue to the next decision making stage Ca Cr  Reject the lot Ca Cr  Accept the lot Cr Cs Reject the lot Cr Cs Accept the lot Cr Cs Continue to the next decision making stage Cr Cs Accept the lot 7. 2.150. 0.9. 4.8.95. 0.1) (100.18]  accept the lot  p  (0. 4.9./ Cost Analysis of Acceptance Sampling.9.1.59072 Reject the lot 83.150. 2.3. 2. 2.2727273 Reject the lot 98. 0. 2. 0.C s .8.95.1. 0. it is found that the optimal policy is a type of control threshold policy.8.05. 2.150.95. 2.18.1. 0.1. 2. 0.1.1) (100.1140686 Reject the lot 94. 0.83957219 Tumble the lot 121.8.1.8. 5. 0.10. 2.9.1. Table 11 Sensitivity analysis N. 0.9.1.43.0.1.150.5. 0. When the nonconforming proportion is more than 0.T . 2.5.95. 0.95.8. 4.25.05. 2.95.S.0. 2. 0.16.95.1. 0.10. Thus.8.52092074 Reject the lot 100.05. 0.45454545 Tumble the lot 141. 2.1336898 Tumble the lot 106.8.1) (100. 0.10.05. 2. 0.2. 2.1. 4.05. 2.150.8.150. 0.1. 0.8395722 Accept the lot 95. then optimal policy is to continue to the next decision making stage therefore 22 . 2. 0.C a . 0. 0.05.150. 4.. 0. 0. 2. 4. 0.9. 0. 0. when the nonconforming proportion is more than 0.10.18.150.1) (100. 0. 4. 0. 2.10.5. m . 0.1) (100.00.88770053 Accept the lot 125.05.150.  .

It is observed that the obtained dynamic model can be solved recursively using a heuristic method.33 Continue 5 8 0.27 Continue optimal policy Reject 4 a 5 12 b 11 0.38 Continue 5 9 0.29 Continue optimal policy Continue 3 a 4 12 b 11 0.13 Accept 3 13 0. Conclusions In this paper.21 3 9 0.43 Reject the cases that inspection is perfect and the second one considers inspection errors in the model. we developed two optimization models for acceptance sampling plan.10 p 0.40 Continue 4 7 0.36 Continue 4 8 0.31 Continue 4 10 0.18 p 0.40 Continue 7 9 0.25 Continue 4 9 0.17 Accept optimal policy   p 3 10 0.08 Accept 2 11 0.27 Accept optimal policy Continue   p 2 10 0.31 Continue optimal policy Continue   p optimal policy   p optimal policy   p optimal policy   p optimal policy   p optimal policy   p optimal policy   p optimal policy 4 6 0.2 p 0.36 Continue 5 10 0.37 Continue   p optimal policy   p optimal policy 6 9 0.14 Accept 2 13 0.09 Accept 1 11 0.18 Accept   p optimal policy   p optimal policy   p optimal policy   p optimal policy   p optimal policy   p optimal policy   p 3 7 0.25 p 0.15 Accept 2 12 0.33 Continue 6 10 0.20 Accept optimal policy 1 10 0.Journal of Optimization in Industrial Engineering 19 (2016) 9-24 Table 13 Sensitivity analysis of “  and  ” m=9   p optimal policy m=10   p optimal policy m=11   p optimal policy m=12   p optimal policy m=13   p optimal policy m=14   p optimal policy m=15   p optimal policy m=16   p optimal policy 1 8 0. To achieve this goal. The first model is written for 23 .23 Continue optimal policy   p 3 11 0.11 Accept 1 a 2 9 b 8 0.3 Continue optimal policy Continue 2 a 3 9 b 8 0. we used a dynamic programming model and 8.

.. 1099-1106.. Quality and Reliability Engineering International. Bayesian inference to design a sampling plan in quality environments considering quality costs.B. 16-28. 27(7). (2012). rejecting and inspection all items. Lindley. Li. M. 25(1). International Journal of Engineering.. Biometrika. first the probability distribution of nonconforming proportion is obtained by Bayesian inference then the correct value of nonconforming proportion is obtained by considering inspection errors then.. S.S. 40(4).B. M.S.S.. Integrated Maintenance and Production Control of a Deteriorating Production System..S. 702-725. S.. Barnett. 38. M. Niaki. P. 34(3). Klassen.S. IIE Transactions. Iranian Journal of Operations Research.. Nair. 43.K..S. Fallahnezhad.A. B.. then the optimal decision is to continue to the next decision making stage. Fallahnezhad. 37(9). Journal of Quality Technology. A new Markov Chain Based Acceptance Sampling Policy via the Minimum Angle Method. (2001). Designing a Single Stage Acceptance Sampling Plan based on the control Threshold policy. Communications in Statistics . A new economical design of acceptance sampling models using Bayesian inference. 45-54. 43. and when the probability of first type error increases. S. then the optimal policy tends to reject the lot. (2012).S.. H.. S. London.T. A new acceptance sampling design using Bayesian modeling and backwards induction. 7.. 3618-3627. 171. E.A. Academic Press. 187-195. we considered a cost function and tried to optimize the function. it is observed that. B. By comparing the first model and the second model. Fallahnezhad. M. (2015).A. 16(1).. A multi-stage twomachine replacement strategy using mixture models. 52. Hosseini Nasab. 51 .T. (2013). A backward recursive inference for optimal sequential sampling plans.. A New Acceptance Sampling Policy Based on Number of Successive Conforming Items. Nembhard. (1965).A. (1992). Nav Res Logist Q. (1983). 19(6). Accred Qual Assur. A modeling approach to maintenance decisions using statistical quality control and optimization. H. to appear in Communications in Statistics-Simulation and Computation. European Journal of Operational Research..T.. when the lot size decreases. and Chang. Owlia. (1971). Sequential sampling: Two decision problems with linear losses for binomial and normal random variables. Pattern Recognition. Fallahnezhad. 1542-1552.S. 198-207.. S. Scientia Iranica. Fallahnezhad. Duenyas. Tang. Furthermore. It is observed that when the lot size is less than a threshold. 3(1). 355-366. (2013).. C. Communications in Statistics-Theory and Methods.. A Finite Horizon Dynamic Programming Model for Production and Repair Decisions.. M.A.. Credit in acceptance sampling on attributes.S. M. Niaki... A New Acceptance Sampling Plan Using Bayesian approach in the presence of inspection errors. (2005). T.A. Hosseini Nasab. 212-222.. Yousefi Babadi. Also when the cost of tumbling process is less than a threshold. Technometrics. 203–213. 18. Journal of Testing and Evaluation. Transactions of the Institute of Measurement and Control. Niaki. A. H.N. Ivy.. Communications in Statistics—Theory and Methods. (2011)... S. Kuo.S. D.. References Aslam...58.S. International Journal of Science and Technology (Scientia Iranica). Aslam. Iravani.S. although the cost of the second model (considering inspection errors) is more than the cost of the first model (perfect inspection)..1073. U. M. M. (2014).. 3302–3313. 5(4).. (2009).A. 340-347.S.. Bayesian variables acceptance-sampling plans: Quadratic loss function and step loss function. Fallahnezhad. S. M. Sadeghieh A.S. then the optimal decision is to tumble the lot. A./ Cost Analysis of Acceptance Sampling. Fallahnezhad. 24 . Vahdat Zad. (2005). Fallahnezhad. H.. Fallahnezhad. A. Designing an optimum acceptance plan using Bayesian inference and stochastic dynamic Programming. K..F. M. Mostafaei.. (2001). and Niaki S. Yousefi A.. Niaki. 507-532.A. Fallahnezhad.T. An iterative decision rule to minimize cost of acceptance sampling plan in machine replacement problem. tumbling the lot or continuing and more sampling in the presence of inspection errors. in the second model. 826-832. 34 . 19-25.. J. Optimal adaptive control policy for joint machine maintenance and product quality control. M. Introduction to Stochastic Dynamic Programming. 104111. and I. L.T. V... Ross. A recursive Approach for lot Sentencing problem in the Presence of inspection errors. Wortham. M. (2016). M. Y. A New Bayesian Acceptance Sampling Plan with Considering Inspection Errors.Theory and Methods. (2012). Sajjadieh. (2013).. 1060. (2007). Fallahnezhad MS. M. then the optimal policy is completely different in some cases. A Multi-Stage Single-Machine Replacement Strategy Using Stochastic Dynamic Programming. and when the lot size is more than this threshold.V... Decision Procedure for the Weibull Distribution based on Run Lengths of Conforming Items. S. Moskowitz. Bayesian inference for some mixture problems in quality and reliability. Technometrics. then the optimal decision is to tumble the lot. 21. 33. M. Niaki. International Journal of Engineering. 423-435. (2002). N. Bayesian inference and stochastic dynamic programming. Fallahnezhad. Xu. M. International Journal of Industrial Engineering & Production Research.W. Fallahnezhad. In order to determine the optimal policy. it is observed that the optimal policy is a type of control threshold policy. 18.J. Fallahnezhad.S. Classification on nonconforming items using unidentified samples. Abdollahi.. M. Tang. Azam.. 41(5). (2011). (2006). 143-150. Using sensitivity analysis.. Fallahnezhad. M. A decision making framework in production processes using Bayesian inference and stochastic dynamic programming. (2014). Journal of Industrial and Systems Engineering.. 586-597. the dynamic programming approach is utilized to make the optimal decision among decisions of accepting. M. 22(3). M. 42(8). Journal of Applied Science. (2012).T.Mohammad Saber Fallah Nezhad et al. Wilson. when the inspections are not perfect. 1865-1869.S. then the optimal decision tends to accept the lot.