Reliability Engineering and System Safety 66 (1999) 49–55

www.elsevier.com/locate/ress

Economic analysis of software release problems with warranty cost and

reliability requirement
M. Kimura*, T. Toyota, S. Yamada
Department of Social Systems Engineering, Faculty of Engineering, Tottori University, Tottori-shi, 680-8552 Japan
Received 28 March 1998; accepted 7 February 1999

Abstract
We discuss optimal software release problems which consider both a present value and a warranty period (in the operational phase) during
which the developer has to pay the cost for fixing any faults detected. It is very important with respect to software development management
that we solve an optimal software testing time by integrating the total expected testing cost and the reliability requirement. We apply a
nonhomogeneous Poisson process model to the formulation of a software cost model and analyze three typical cases of the cost model.
Moreover, we derive several optimal release polices. Finally, numerical examples are shown to illustrate the results of the optimal policies.
q 1999 Elsevier Science Ltd. All rights reserved.
Keywords: Optimal software release problem; Software reliability growth model; Nonhomogeneous Poisson process; Warranty cost

1. Introduction increases. However, it leads to increase the testing cost

and to delay software delivery. In contrast, if the length of
Today, computer systems are indispensable in our lives, software testing is short, a software system with low relia-
and their importance and needs have been increasing. Thus, bility is delivered and it includes many software faults
we require more large-scale and complex software systems which have not been removed in the testing phase. Thus,
to construct such computer systems. However, a computer the maintenance cost during the operation phase increases.
software is developed by human work, so many software So, it is very important in terms of software management
faults must be introduced into the software product during that we solve for the optimal length of software testing, what
its development process. Therefore, it is very important to is called an optimum release time. Such a decision problem
apply useful technologies for developing a highly reliable is called an optimal software release problem [1–8]. These
software system. decision problems have been studied in the last decade by
Recently, it is becoming increasingly difficult for the many researchers. Several optimal policies for releasing the
developers to produce highly-reliable software systems effi- software products have been actually used in the real soft-
ciently. Thus, it has been necessary to control a software ware development processes. For instance, Chaar et al. [9]
development process in terms of reliability, cost, and release have treated the issues in terms of the software inspection
time. In the last phase of software development process, the and test, and have discussed them minutely by introducing
testing is carried out to detect and fix software faults intro- the concept of defect trigger during the actual software
duced by human work, prior to its release for the operational development processes.
use. The software faults that cannot be detected and fixed In this article, we discuss several optimal release policies
remain in the released software system after the testing which provide the total testing time minimizing the total
phase. Thus, if a software failure occurs during the opera- expected software cost and satisfying a reliability require-
tional phase, then a computer system stops working and it ment. Moreover, it is assumed that there will be a warranty
may cause serious damage to our daily life. period [10] at the beginning of the operational phase during
If the length of software testing is long, we can remove which the developer has to pay the cost for fixing any faults
many software faults in the system and its reliability detected. We also introduce the concept of present value
[3,6] into the cost factors.
In the following, we describe the software reliability
* Corresponding author. Fax: 1 81-857-31-0882.
E-mail addresses: [email protected] (M. Kimura); [email protected] growth models (SRGMs) [11–16] based on a non-
tottori-u.ac.jp (S. Yamada) homogeneous Poisson process (NHPP) and the total
0951-8320/99/$ - see front matter q 1999 Elsevier Science Ltd. All rights reserved.
PII: S0951-832 0(99)00020-4
50 M. Kimura et al. / Reliability Engineering and System Safety 66 (1999) 49–55

{m t†}n
Pr{N t† ˆ n} ˆ exp‰2m t†Š n ˆ 0; 1; 2; …†;
n!
1†
Zt
m t† ˆ hm t 0 †dt 0 ; 2†
0

where Pr{A} means the probability of event A. In Eqs. (1)

and (2), m(t) is called a mean value function which repre-
sents the expected cumulative number of faults detected in
the testing time interval (0,t], and hm t† an intensity function
which denotes the fault detection rate per unit time.
In this article, we apply an exponential software reliabil-
Fig. 1. Software reliability growth aspects during the warranty period.
ity growth model which is based on an NHPP to the software
fault-detection phenomenon. Its mean value and intensity
functions are given as follows:
expected maintenance cost is formulated in Section 2. The m t† ˆ a 1 2 e2bt † a . 0; b . 0†; 3†
optimal software release problems with cost and reliability
requirements are represented in Section 3. We analyze the hm t† ˆ abe2bt ; 4†
software maintenance cost model and derive the optimal
release policies in Section 4. Numerical examples are where the parameters a and b are the expected number of
shown to illustrate the optimal policies in Section 5. initial fault content and the fault detection rate per one fault,
respectively.
This model is often applied to a software failure-occur-
rence phenomenon for a large-scale software system, in
2. Applied SRGM which the testing time is measured by the CPU time.

In the testing phase of software development and its

3. Software maintenance cost models
operational phase, it is natural that we should apply prob-
ability and statistic theories to a software failure-occurrence The following notations are defined:
or software fault-detection phenomenon for describing the
stochastic behavior. Generally, the implemented software c0 initial testing cost which is the barest minimum
system is investigated whether the system can perform prop- requirement;
erly by its specification requirements. Concretely, many ct testing cost per unit time;
test-cases which are generated prior to the testing phase cw maintenance cost per one fault during the warranty
from the specification requirements, are executed one by period;
one in the (system) testing procedure. Each execution result T software release time, i.e. additional total testing
is confirmed whether it is acceptable from the viewpoint of time;
the specification. If the result is unacceptable, the software T* optimum software release time.
has at least one software fault, then the fault must be elimi-
nated. From the viewpoint of such a black-box testing 3.1. Optimal software release problem
process, software failure-occurrence and fault-detection
phenomena can be treated as stochastic processes. These We discuss the formulation of the optimal release
phenomena can be formulated by various mathematical problem. The maintenance cost due to the warranty period
models, which are called software reliability models. Espe- is considered. The concept of a present value is also intro-
cially, as the faults latent in the software system are detected duced into the cost factors. Then, the total expected software
and fixed by the testing, if it is assumed that no new faults maintenance cost WC(T) can be formulated as:
are introduced into the system, the probability of software ZT
failure-occurrence decreases, i.e. the software reliability WC T† ; c0 1 ct e2at dt 1 Cw T†; 5†
0
increases. In this article, we apply an NHPP to describe
the time-dependent behavior of the cumulative number of where Cw(T) is the maintenance cost during the warranty
faults detected up to a certain testing time. period. The parameter a in Eq. (5) is a discount rate of the
Let N(t) and m(t) be the cumulative number of faults cost. In this article, we discuss the following three cases in
detected up to testing time t and the expectation of N(t), terms of the behavior of Cw(T) (see Fig. 1):
respectively. Then, the software fault detection phenom- Case 1: When the length of the warranty period is
enon can be represented as an NHPP: constant and the software reliability growth is not assumed
 M. Kimura et al. / Reliability Engineering and System Safety 66 (1999) 49–55 51

to occur after the testing phase, Cw(T) is represented as: 4. Optimal software release policies
ZT 1 Tw
Cw T† ˆ cw hm T†e2at dt: 6† 4.1. Analysis of maintenance cost models
T
We discuss the optimum release time minimizing the
Case 2: When the length of the warranty period is total expected software cost WC(T) in Eq. (5) based on
constant and the software reliability growth is assumed to the exponential software reliability growth model formu-
occur even after the testing, Cw(T) is given by: lated by Eqs. (3) and (4) in respective cases.
ZT 1 Tw First, we consider the optimal release policies of Case 1.
Cw t† ˆ cw hm t†e2at dt: 7† Substituting Eq. (6) into Eq. (5), we rewrite it as:
T
ZT ZT 1 Tw
Case 3: When the length of the warranty period obeys a WC T† ˆ c0 1 ct e2at dt 1 cw hm T† e2at dt: 13†
0 T
distribution function W(t) and the software reliability
growth is assumed to occur even after the testing phase, Differentiating Eq. (13) in terms of T and equating it to zero
Cw(T) is represented as: yields:
Z∞ ZT 1 Tw ct
hm t† ˆ : 14†
Cw t† ˆ cw hm t†e2at dtdW Tw †: 8† cw Tw b 1 a†
0 T
Note that WC(T) is a convex function with respect to T
where we assume that the distribution of the warranty period because d 2WC(T)/dT 2 . 0. Thus, the equation dWC(T)/
is a truncated normal distribution: dT ˆ 0 has only one finite solution when the condition
dW t† 1 hm 0† . ct =‰cw Tw b 1 a†Š holds. The solution T1 of Eq.
ˆ p e{2 t2m† =2s }
2 2
t $ 0; m . 0; s . 0†; (14) and the optimum release time can be shown as follows:
dt A 2ps  
9† 1 abcw Tw b 1 a†
T* ˆ T1 ˆ ln 0 , T1 , ∞†: 15†
b ct
1 Z∞ {2 t2m†2 =2s2 } When the condition holds, WC(T) in Eq. (13) is a mono-
A ˆ p e dt:
2p s 0 tonically increasing function in terms of the testing time T.
Then, the optimum release time is T* ˆ 0 Therefore, we
obtain the optimal release policies as follows:
3.2. Simultaneous cost and reliability requirements [Optimal Release Policy 1]

Next, we discuss the optimal release problem with the (1.1) If hm 0† . ct =‰cw Tw b 1 a†Š; then the optimum
requirement of attained software reliability. In the actual release time is T* ˆ T1 .
software development, the manager must spend and control (1.2) If hm 0† # ct =‰cw Tw b 1 a†Š, then the optimum
the testing resources as minimizing the total software cost release time is T* ˆ 0.
and satisfying reliability requirement rather than only mini- Next, we derive the optimal release policies of Case 2.
mizing the cost. From the NHPP model formulated by Eqs. Substituting Eq. (7) into Eq. (5), we obtain WC(T) as:
(3) and (4), the software reliability function can be defined ZT ZT 1 Tw
as the probability that a software failure does not occur WC T† ˆ c0 1 ct e2at dt 1 cw hm t†e2at dt: 16†
during the time-interval T; T 1 xŠ after the testing time T, 0 T
i.e. the release time. The software reliability function is Differentiating Eq. (16) in terms of T and equating it to zero
given as follows: yields:
R xuT† ˆ exp‰2{m T 1 x† 2 m T†}Š: 10† ct
hm T† ˆ T . 0†: 17†
cw ‰1 2 e2 b1a†Tw Š
Substituting Eq. (3) into Eq. (10), we derive the software
reliability function as follows: By using the same procedure as Case 1, we have the finite
solution of Eq. (17) and the optimum release time as
R xuT† ˆ exp‰2e2bt ·m x†Š: 11† follows:
( )
Let the software reliability objective be R0 0 , R0 # 1†: 1 abcw ‰1 2 e2 b1a†Tw Š
T* ˆ T2 ˆ ln 0 , T2 , ∞†;
We can evaluate optimum release time T ˆ T* which mini- b ct
mizes Eq. (5) with satisfying the software reliability objec- 18†
tive R0. Thus, the optimal software release problem is 2 b1a†Tw
formulated as follows: when hm 0† . ct ={cw ‰1 2 e Š} holds. When the
condition hm 0† # ct ={cw ‰1 2 e2 b1a†Tw Š} holds, WC(T) in
minimize WC T† subject to R xuT† $ R0 : 12† Eq. (16) is a monotonically increasing function in terms of
52 M. Kimura et al. / Reliability Engineering and System Safety 66 (1999) 49–55

the testing time T. Then, the optimum release time is let TR be the optimum release time with respect to T satisfy-
T* ˆ 0. Therefore, we obtain the optimal release policy as ing the relation R xuT† ˆ R0 . By applying the relation
follows: R xuT† ˆ R0 into Eq. (11), we can obtain the solution TR
[Optimal Release Policy 2] as follows:
  
(2.1) If hm 0† . ct ={cw ‰1 2 e2 b1a†Tw Š}, then the opti- 1 1
TR ˆ ln m x† 2 ln ln : 24†
mum release time is T* ˆ T2 . b R0
(2.2) If hm 0† # ct ={cw ‰1 2 e2 b1a†Tw Š}, then the opti-
mum release time is T* ˆ 0. Then, we can derive the optimal release policies to mini-
mize the total expected software maintenance cost and to
Finally, we obtain the optimal release policies of Case 3 satisfy the software reliability objective R0.
as follows. Substituting Eq. (8) into Eq. (5), we can derive First, for Case 1, the optimal release policies are given as
Eq. (5) as: follows:
ZT [Optimal Release Policy 4]
WC T† ˆ c0 1 ct e2at dt
0 (4.1) If hm 0† . ct =‰cw Tw b 1 a†Š and R xj0† , R0 , then
Z∞ ZT 1 Tw the optimum release time is T* ˆ max{T1 ; TR }.
1 cw hm t†e2at dtdW Tw †: 19† (4.2) If hm 0† . ct =‰cw Tw b 1 a†Š and R xu0† $ R0 , then
0 T the optimum release time is T* ˆ T1 .
Differentiating Eq. (19) in terms of T and equating it to zero (4.3) If hm 0† # ct =‰cw Tw b 1 a†Š and R xu0† , R0 , then
yields: the optimum release time is T* ˆ TR .
c (4.4) If hm 0† # ct =‰cw Tw b 1 a†Š and R xu0† $ R0 , then
hm T† ˆ t T . 0†; 20† the optimum release time is T* ˆ 0.
cw g
Second, for Case 2, the optimal release policies are given as
where
follows:
gˆ1 [Optimal Release Policy 5]
! (5.1) If hm 0† . ct ={cw ‰1 2 e2 b1a†Tw Š} and
m 2 s2 a 1 b†
F " # R xu0† , R0 , then the optimum release time is
s 22m a 1 b† 1 a 1 b†2 s2 T* ˆ max{T2 ; TR }.
2   exp ;
m 2 (5.2) If hm 0† . ct ={cw ‰1 2 e2 b1a†Tw Š} and
F
s R xu0† $ R0 , then the optimum release time is
21† T* ˆ T2 .
(5.3) If hm 0† # ct ={cw ‰1 2 e2 b1a†Tw Š} and
and
R xu0† , R0 , then the optimum release time is
1 Zx T* ˆ TR .
e2 y =2† dy:
2
F x† ˆ p 22†
2p 2 ∞ (5.4) If hm 0† # ct ={cw ‰1 2 e2 b1a†Tw Š} and
R xu0† $ R0 , then the optimum release time is
Note that the equation dWC(T)/dT ˆ 0 has only one finite
T* ˆ 0.
solution when the relation hm 0† . ct = cw g† holds. The opti-
mum release time is: Third, for Case 3, the optimal release policies are given as
  follows:
1 abcw g
T* ˆ T3 ˆ ln 0 , T3 , ∞†; 23† [Optimal Release Policy 6]
b ct
(1) If hm 0† . ct = cw g† and R xu0† , R0 , then the
In contrast, when hm 0† # ct = cw g†, the optimum release optimum release time is T* ˆ max{T3 ; TR }.
time is T* ˆ 0. The optimal release policies for Case 3 (2) If hm 0† . ct = cw g† and R xu0† $ R0 , then the
are given as follows: optimum release time is T* ˆ T3 .
[Optimal Release Policy 3] (3) If hm 0† # ct = cw g† and R xu0† , R0 , then the
(3.1) If hm 0† . ct = cw g†, then the optimum release time optimum release time is T* ˆ TR .
is T* ˆ T3 . (4) If hm 0† # ct = cw g† and R xu0† $ R0 , then the
(3.2) If hm 0† # ct = cw g†, then the optimum release time optimum release time is T* ˆ 0.
is T* ˆ 0.

4.2. Analysis of maintenance cost models with a reliability 5. Numerical examples

requirement
In this section, we show numerical examples to illustrate
For the optimal release problem formulated by Eq. (12), the optimal release policies derived in the previous section.
 M. Kimura et al. / Reliability Engineering and System Safety 66 (1999) 49–55 53

Table 1 Table 3
Dependence of Tw and ct, or, m and ct on the optimum release time (c0 ˆ Dependence of R0 and x on the optimum release time without cost factor
1000.0, cw ˆ 20.0, a ˆ 1000, b ˆ 0.05, m /s ˆ 10, a ˆ 0.001) consideration (a ˆ 1000, b ˆ 0.05)

Tw ct R0 x

1.0 3.0 5.0 10.0 30.0 50.0 1.0 3.0 5.0 10.0 50.0

(Case 1) [Optimal Release Policy 1] 0.95 137.1 158.1 167.4 178.9 195.8
1.0 78.6 56.7 46.4 32.6 10.6 0.4 0.90 122.8 143.7 153.0 164.5 181.4
2.0 92.5 70.5 60.3 46.4 24.5 14.3 0.85 114.1 135.1 144.3 155.8 172.8
5.0 110.8 88.9 78.7 64.8 42.8 32.6 0.80 107.7 128.7 138.0 149.5 166.4
10.0 124.7 102.7 92.5 78.6 56.7 46.4 0.75 102.7 123.7 132.9 144.4 161.4
20.0 138.6 116.6 106.4 92.5 70.5 60.3 0.70 98.4 119.4 128.6 140.1 157.1
50.0 156.9 134.9 124.7 110.8 88.9 78.6 0.65 94.6 115.6 124.8 136.3 153.3
100.0 170.7 148.8 138.6 124.7 102.7 92.5 0.60 91.2 112.2 121.4 132.9 149.9
0.55 88.0 109.0 118.3 129.8 146.7
(Case 2) [Optimal Release Policy 2] 0.50 85.1 106.1 115.3 126.8 143.8
1.0 78.1 56.2 45.9 32.1 10.1 0
2.0 91.5 69.5 59.3 45.4 23.5 13.2
5.0 108.3 86.3 76.1 62.3 40.3 30.1
10.0 119.8 97.8 87.6 73.8 51.8 41.6
We assume the parameters in the exponential SRGM and
20.0 129.2 107.2 97.0 83.2 61.2 51.0 cost factors as follows:
50.0 136.5 114.6 104.3 90.5 68.5 58.3
100.0 138.0 116.1 105.8 92.0 70.0 59.8 c0 ˆ 1000:0; cw ˆ 20:0; a ˆ 1000:0; b ˆ 0:05:
(Case 3) [Optimal Release Policy 3] Table 1 shows the optimum release times for several values
1.0 78.3 56.3 46.1 32.3 10.3 0.1
2.0 91.7 69.7 59.5 45.6 23.6 13.4
of the testing ct and the warranty period Tw or m . As the
5.0 108.5 86.5 76.2 62.4 40.5 30.2 length of the warranty period is long, the optimum release
10.0 120.0 97.9 87.7 73.9 51.9 41.7 time increases. Thus, it is shown that we have to spend
20.0 129.3 107.3 97.1 83.2 61.2 51.0 enough time to test the software system when the warranty
50.0 136.5 114.6 104.3 90.5 68.5 58.3 period is long. As the testing cost per unit time increases, the
100.0 138.0 116.1 105.8 92.0 70.0 59.8
Table 4
Table 2 Dependence of Tw and ct, or, m and ct on the optimum release time (c0 ˆ
Dependence of Tw and a , or, m and a on the optimum release time (c0 ˆ 1000.0, cw ˆ 20.0, a ˆ 1000, b ˆ 0.05, m /s ˆ 10, a ˆ 0.001, x ˆ 1.0, R0 ˆ
1000.0, cw ˆ 20.0, ct ˆ 5.0, a ˆ 1000, b ˆ 0.05, m /s ˆ 10) 0.8)

Tw a Tw ct

0.0 0.001 0.005 0.01 0.05 0.1 1.0 3.0 5.0 10.0 30.0 50.0

(Case 1) [Optimal Release Policy 1] (Case 4) [Optimal Release Policy 4]

1.0 46.1 46.4 48.0 49.7 59.9 68.0 1.0 107.7 107.7 107.7 107.7 107.7 107.7
2.0 59.9 60.3 61.8 63.6 73.8 81.9 2.0 107.7 107.7 107.7 107.7 107.7 107.7
5.0 78.2 78.6 80.1 81.9 92.1 100.2 5.0 110.8 107.7 107.7 107.7 107.7 107.7
10.0 92.1 92.5 94.0 95.7 106.0 114.1 10.0 124.7 107.7 107.7 107.7 107.7 107.7
20.0 106.0 106.4 107.9 109.6 119.8 127.9 20.0 138.6 116.6 107.7 107.7 107.7 107.7
50.0 124.3 124.9 126.2 127.9 138.2 146.3 50.0 156.9 134.9 124.7 110.8 107.7 107.7
100.0 138.2 138.6 140.1 141.8 152.0 160.1 100.0 170.7 148.8 138.6 124.7 102.7 107.7

(Case 2) [Optimal Release Policy 2] (Case 2) [Optimal Release Policy 5]

1.0 45.6 45.9 47.4 49.1 58.9 66.5 1.0 107.7 107.7 107.7 107.7 107.7 107.7
2.0 58.9 59.3 60.7 62.4 71.8 79.0 2.0 107.7 107.7 107.7 107.7 107.7 107.7
5.0 75.8 76.1 77.5 79.0 87.3 93.2 5.0 108.3 107.7 107.7 107.7 107.7 107.7
10.0 87.3 87.6 88.8 90.0 96.8 100.9 10.0 119.8 107.7 107.7 107.7 107.7 107.7
20.0 96.8 97.0 97.9 98.8 103.1 104.9 20.0 129.2 107.7 107.7 107.7 107.7 107.7
50.0 104.3 104.3 104.6 104.9 105.8 106.0 50.0 136.5 114.6 107.7 107.7 107.7 107.7
100.0 105.8 105.8 105.9 106.0 106.0 106.0 100.0 138.0 116.1 107.7 107.7 107.7 107.7

(Case 3) [Optimal Release Policy 3] (Case 6) [Optimal Release Policy 6]

1.0 45.5 45.9 47.4 49.1 58.9 66.5 1.0 107.7 107.7 107.7 107.7 107.7 107.7
2.0 58.9 59.3 60.7 62.4 71.8 78.9 2.0 107.7 107.7 107.7 107.7 107.7 107.7
5.0 75.8 76.1 77.4 78.9 87.3 93.1 5.0 108.5 107.7 107.7 107.7 107.7 107.7
10.0 87.3 87.6 88.7 90.0 96.7 100.9 10.0 120.0 107.7 107.7 107.7 107.7 107.7
20.0 96.7 97.0 97.8 98.7 103.1 104.9 20.0 129.3 107.7 107.7 107.7 107.7 107.7
50.0 104.3 104.3 104.6 104.9 105.8 106.0 50.0 136.5 114.6 107.7 107.7 107.7 107.7
100.0 105.8 105.8 105.9 105.9 106.0 106.0 100.0 138.0 116.1 107.7 107.7 107.7 107.7
54 M. Kimura et al. / Reliability Engineering and System Safety 66 (1999) 49–55

Fig. 2. The optimum software release time with cost and reliability requirement. (c0 ˆ 1000.0, ct ˆ 5.0, cw ˆ 20.0, a ˆ 1000, b ˆ 0.05, a ˆ 0.001, x ˆ 1.0, Tw ˆ
10).

optimum release time decreases. When we compare Case 1 Table 4 shows the optimum release time for several
with Case 2, the optimum release time of Case 2 is always values of ct and Tw or m , where we assume the values of
shorter than that of Case 1. The reason is that the software the software reliability objective and the operational time
reliability growth occurs after the release. In Case 2 in are 0.8 and 1.0, respectively. Though Table 4 shows the
comparison with Case 3, the optimum release times are same trend as Table 1, the optimum release time satisfying
equivalent. As we assume the value of m /s is 10, the the software reliability objective is a minimum of the values
shape of the distribution of the warranty period which in respective tables.
obeys a truncated normal distribution approximates that of Fig. 2 illustrates the optimum release time of Case 1 with
a normal distribution N m; s2 †. The more the value of m reliability objective R0 for [Optimal Release Policy 4]. In
increases, the stronger the trend becomes. the case of ct ˆ 5.0, a ˆ 0.001 and Tw ˆ 10, we derive T1 ˆ
Table 2 shows the optimum release times with the 92.5 from [Optimal Release Policy 1] of Table 1. When the
discount rate a for several values of the warranty period reliability objective and operational time are assumed to be
Tw or m . Amongst Case 1, Case 2 and Case 3, the longer the R0 ˆ 0.9 and x ˆ 1.0, we obtain TR ˆ 122.8 in Table 3.
warranty period is, the more the optimal release time Therefore, from [Optimal Release Policy 4] the optimum
increases. As the discount rate increases, the optimum release time is decided as follows:
release time increases.
T* ˆ max{T1 ; TR } ˆ max{92:5; 122:8} ˆ 122:8
Now, we show the optimum release time based on only
reliability criterion without cost factor consideration before weeks†:
discussing that with both cost factor and reliability consid-
erations. Table 3 shows the optimum release time satisfying
a reliability objective in case of the variation of the software 6. Conclusion
reliability objective R0 and the operational time x. The
longer the operational time is, the more the optimum release In this article, we have discussed the method of determin-
time increases. ing an optimum release time by using the maintenance cost
 M. Kimura et al. / Reliability Engineering and System Safety 66 (1999) 49–55 55

model. Software cost factors and software reliability cycle and delivery delay. In: Proc. 2nd ISSAT Int. Conf. Reliability
requirement in the testing/operational phase have been and Quality in Design, 1995, pp. 215–219.
[4] Koch HS, Kubat P. Optimal release time for computer software. IEEE
considered. Especially, we have introduced the concept of Trans. Softw. Engng 1983;SE-9(3):323–327.
present value into the cost factors and have considered [5] Okumoto K, Goel AL. Optimum release time for software system
warranty period in the operational phase. We have applied based on reliability and cost criteria. J. Systems and Software
the exponential SRGM to the formulation of the total 1980;1(4):315–318.
expected software cost and considered three cases of the [6] Yamada S. Optimal release problems with warranty period based on a
software maintenance cost model. Trans. IPS Japan
maintenance cost during the warranty period. Then, the 1994;35(10):2197–2202 in Japanese.
optimal release policies have been derived and numerical [7] Yamada S, Kimura M, Teraue E, Osaki S. Optimal software release
examples based on them shown. problems with life-cycle distribution and discount rate. Trans. IPS
In the future, it is necessary to examine the optimal Japan 1993;34(5):1188–1197 in Japanese.
release policies by applying to real software projects. [8] Yamada S, Osaki S. Optimal software release policies with simulta-
neous cost and reliability requirements. Eur. J. Oper. Res.
1987;31(1):46–51.
[9] Chaar JK, Halliday MJ, Bhandari IS, Chillarege R. In-process evalua-
Acknowledgements tion for software inspection and test. IEEE Trans. Softw. Engng
1993;19(11):1055–1070.
This work was supported in part by a Grant-in-Aid for [10] Malaiya YK, Srimani PK, editors. Software reliability models: theo-
Scientific Research from the Ministry of Education, retical developments, evaluation and applications Los Alamitos:
Science, and Culture of Japan (Grant No. 10680431, IEEE Computer Society Press, 1990.
[11] Musa JD, Iannino A, Okumoto K. Software reliability: measurement,
09780410). prediction, application. New York: McGraw-Hill, 1987.
[12] Pham H, editor. Software reliability and testing Los Alamitos: IEEE
Computer Society Press, 1995.
References [13] Yamada S. Software quality/reliability measurement and assessment:
Software reliability growth models and data analysis. J. Info. Process.
[1] Cho BC, Park KS. An optimal time for software testing under the 1991;14(3):254–266.
user’s requirement of failure-free demonstration before release. [14] Xie M. Software reliability modelling. Singapore: World Scientific,
IEICE Trans. Fundamentals 1994;E77-A(3):563–570. 1991.
[2] Foreman EH, Singpurwalla ND. Optimal time intervals for testing- [15] Yamada S, Ohtera H. Software reliability: theory and practical appli-
hypotheses on computer software errors. IEEE Trans. Reliability cations (in Japanese), Soft Research Center, Tokyo, 1990.
1979;R-28(3):250–253. [16] Shooman ML. Software engineering: design, reliability, and manage-
[3] Kimura M, Yamada S. Optimal software policies with random life- ment. New York: McGraw-Hill, 1983.