Table of Contents
ANALYSIS of STOCK MANAGEMENT GAMING EXPERIMENTS
and ALTERNATIVE ORDERING FORMULATIONS’
Yaman Barlas Mehmet Giinhan Ozevin
Department of Industrial Engineering Ford Otomotiv Sanayi A.S.
Bogazici University izmit Gélciik Yolu 14. Km., ihsaniye
80815, Bebek, Istanbul, TURKEY 41670, Gélciik, Kocaeli, TURKEY
Tel:+90 (212) 359 70 73 Tel:+90 (262) 315 53 25
ybarlas@boun.edu.tr gozevin@ford.com.tr
Abstract
This paper investigates two different yet related research questions about stock
management in feedback environments: The first one is to analyze the effects of
selected experimental factors on the performances of subjects (players) in a stock
management simulation game. In light of these results, our second objective is to
evaluate the adequacy of standard decision rules typically used in dynamic stock
management models and to seek improvement formulations. To carry out the research,
the generic stock management problem is chosen as the interactive gaming platform. In
the first part, gaming experiments are designed to test the effects of three factors on
decision making behavior: different patterns of customer demand, minimum possible
order decision (‘review’) interval and finally the type of the receiving delay. ANOVA
results of these 3-factor, 2-level experiments show which factors have significant effects
on ten different measures of behavior (such as max-min range of orders, inventory
amplitudes, periods of oscillations and backlog durations). In the second phase of
research, the performances of subjects are compared against some selected ordering
heuristics (formulations). First, the patterns of ordering behavior of subjects are
classified into three basic types. Comparing these three patterns with the stand-alone
simulation results, we observe that the common linear "Anchoring and Adjustment
Rule." can mimic well the smooth and gradually damping type of behavior, but can not
replicate the non-linear and/or discrete ordering dynamics. Thus, several alternative
non-linear rules are formulated and tested against subjects’ behaviors. Some standard
discrete inventory control rules (such as (s, Q)) common in the inventory management
literature are also formulated and tested. These non-linear and/or discrete rules,
compared to the linear stock adjustment rule, are found to be more representative of the
subjects’ ordering behavior in many cases, in the sense that these rules can generate
nonlinear and/or discrete ordering dynamics. Another major finding is the fact that the
well-documented oscillatory dynamic behavior of the inventory is a quite general result,
not just an artifact of the linear anchor and adjust rule. When the supply line is ignored
or underestimated, large inventory oscillations result, not just with the linear anchor-
and-adjust rule but also with the non-linear rules, as well as the standard inventory
management rules. Furthermore, depending on parameter values, nonlinear ordering
rules are more prone to yield unstable oscillations -even if the supply line is taken into
account.
Keywords: stock management, anchor and adjust heuristic, experimental
testing of decision rules, non-linear decision heuristics, inventory control rules
' Supported by Bogazici University Rereacrh Fund no. 02R102
1. Stock Management Game
For the experimental testing purpose, the generic stock management problem,
one of the most common dynamic decision problems, is chosen as the interactive
gaming environment. The objective of the game is stated as "keeping the inventory level
as low as possible while avoiding any backorders." If there is not enough goods in the
inventory at any time, customer orders are entered as backorders to be supplied later.
"Order decisions” are the only means of controlling the inventory level. The general
structure of the stock management problem is illustrated in Figure 1. (See Sterman
2000). The three empty boxes: Expectation Formation, Goal Formation and Decision
Rule are deliberately left blank, as they are unknown, since they take place in the
“minds” of the players. (Later, in the simulation version of the game, these three boxes
will have to be specified. For instance, the expectation formation will be formulated by
exponential smoothing; inventory goal will be set to inventory coverage times expected
demand and supply line goal will be order delay times expected demand. As for the
decision rule, different formulations will be tried: linear stock adjustment rule, three
different non-linear adjustment rules and finally various standard discrete inventory
control rules.) This notion of “gaming experimentation” to analyze and test subjects’
decision heuristics has been successfully used in system dynamics literature (Sterman
1987, 1989), as well as in experimental psychology (Brehmer 1989).
While playing the game, subjects can monitor the system from information
displays/graphs showing their inventory, supply line levels and customer demand
(Appendix A). Neither the costs associated with high inventories nor costs resulting
from backorders are accounted for explicitly in the simulation game. However, the
relation between keeping these costs as low as possible and the objective of the game is
stated in the instruction given to subjects. In other words, subjects are instructed that
keeping large safety stocks would result in high holding costs but backlogs must also be
avoided, as they would incur large costs due to lost demand. Before beginning the
game, all subjects are given a written instruction presenting the problem and their task
(Appendix B). Time available to accomplish the task is not limited. No explicit, tangible
reward is used to motivate the subjects.
2. Gaming Experiments
The first set of gaming experiments are designed to test the effects of three
factors on decision making behavior of subjects:
(a) Length of order decision (review) interval;
(b) Type of the receiving delay;
(c) Different patterns of customer demand
2.1, Length of Decision Interval
Subjects were allowed to order at “each time unit” in the first group of
experiments (Short Game), whereas they were allowed to order “once every five time
units” in the second group of experiments (Long Game). Short Games are simulated for
100 time units whereas Long Games are simulated for 250 time units and the receiving
delays are also shorter (4 time units) in the short game, longer (10 time units) in the
long game. Thus, the ‘length’ effect is really a package (involving longer receiving
delays and longer game length as well), but this package is summarized by the term
Length of Decision Interval effect, because this particular component will be the focus
in interpreting the experimental results, as will be seen below. It is hypothesized that
subjects being free to make decisions at any point in time versus decisions allowed only
every five time units would influence the difficulty of the game, hence cause differences
in the performances.
oe STOCK ACQUISITION SYSTEM Ae WE) externa rut 6
SUPPLY LINE STOCK
@ eo ad
DECISION WwrLow ouriow |
| 4
rime DELAY
@__decsionpoucyaule A @
PINK NOISE
@_coarorwation A @ b @___EXPECTATION FORMATION A @
FIGURE 1. The Stock Management Problem
2.2. Type of Receiving Delay
The second independent experimental factor is the type of the delay. (The
length of the delay is not an independent factor; since it is changed as an integral part of
the Length of Decision Interval effect described above. Note that if one were to change
the decision interval from one to five days but keep the receiving delay at four, then the
nature of the game would change in an implicit and problematic way, in the sense that
the receiving delay would be four times longer than the order interval in the short game
but shorter than the order interval in the long one). We focus on the type of delay
representation, as different delay types may be appropriate for different inventory
acquisition systems, like continuous exponential delay or discrete delay representations.
Since “Receiving” is the inflow to inventory, its transient behavior may influence
decision-maker’s interpretation of the results of his/her own order decisions. The two
extremes of the exponential delay family, namely the infinite-order discrete delay and
first order exponential delay are chosen as the two levels of the delay factor in the
experimental design. (This is further motivated by a common criticism system dynamics
games that states that continuous delays are not realistic/intuitive, so such games pose
an artificial difficulty for players with no expertise in modeling. So, it would be
interesting to see if subjects’ performances would actually deteriorate in cases involving
continuous delays).
2.3. Patterns of Customer Demand
Until the fifth decision interval, average customer demand remains constant at
20. At the beginning of the fifth decision interval (at time five in the Short Game and at
time 25 in Long Game) an unannounced, one-time increase of 20 units occurs in the
customer demand patterns used in the experiments. Subjects react to the disequilibrium
created by this change. After the step up, demand remains constant at 40 in the first type
of customer demand pattern which we call “step up in customer demand”. (Figure 2).
In the second type of demand pattern, called “step up and down in customer
demand,” a second disturbance, a one-time decrease in demand follows the first
increase, after some time interval, restoring the demand back to its original level of 20
(see Figure 3, where the step-down is set to occur at time 20). The time interval between
the step up and down in customer demand is chosen as roughly half of the natural
periodicity of the model (about 25 days in Short Game and about 60 in Long Game). A
common perception in stock management circles is that ‘poor ordering performance is
caused by complex demand patterns’. The purpose of this particular experimental factor
is to test this claim/hypothesis in a simplified context.
Before the demand patterns described above are used in games, "Pink noise"
(auto-correlated noise) is added to the average patterns to obtain more realistic demand
dynamics. The standard deviation of the white noise is set to 15 percent of average
customer demand. The delay constant of the exponential smoothing (the correlation
time used to create pink noise) is taken as two time units. (See Appendix C for
equations).
To summarize, there are eight combinations of the above three factors across
the two levels of each. So we have a 2° factorial design. Each condition is played six
times (random replications), yielding a total of 48 experiments (see Table 1). Since the
demand pattern is discovered by the subjects once the game is played and because they
can improve their performance by practice, to obtain unbiased results, the same subject
never played two Short Games or two Long Games. However, due to limited number of
subjects, some of the subjects were allowed to play one Short Game and one Long
Game, since transferring experience in-between Short and Long Games is not easy.
TABLE 1. Design of Experiments. (X) indicates the selected level of factors
for the corresponding runs.
Length of Type of Receiving Pattern of Customer
Ordering Interval Delay Demand
Runs Every Every Five Step-Up Step-Up-
Time Time F r and-Down
: e Exponential | Discrete | Customer i
Unit Units Demand Customer
(i day) (5 days) Demand
1 x x x
2 x x x
3 x xX x
4 xX x x
3 xX x x
6 xX x x
7 x x x
8 xX xX x
L 00 + + + ‘
0.00 20 50000 750 10000
Neat Graph pl (Pater 1) Tine 1427 07 May 1998 Per
FIGURE 2. Step Up Customer Demand for Short Games
1 demand
1:
80.004
1: 40.004
4 Wn ry Zw
T
25.00 50.00 75.00 100.00
Née
FIGURE 3. Step Up and Down Customer Demand for Short Games
Graph 1: p2 (Pattern 2) Time 14:29 07 May 1998 Per
2.4. Initial Conditions
Games start at equilibrium. Supply line level is initially set at 80 in Short
Games and 200 in Long Games such that no backordering occurs even when the
decision-maker does not order goods during the first four decision intervals (at the end
of which the disturbance in customer demand creates disequilibrium). Inventory levels
are initially set arbitrarily (at 40 and 200 respectively) so as to satisfy the average initial
customer demand for the first two decision intervals (for two days in Short Games and
ten days in Long Games).
3. Analysis of Experiments
The general, broad behavior pattern of inventory in majority of games is one
of oscillations. (See Figures 4, 5 and most other games illustrated in the Figures through
the article). This finding is consistent with overwhelming evidence on oscillating
inventories in system dynamics literature and elsewhere (for instance, Forrester 1961,
Sterman 1989, 2000, Lee et al 1997 and Tvede 1996). We will return later to this main
‘qualitative’ result. But first, we take a more quantitative look at the effects of the three
factors.
Representative summary measures of orders and inventory levels computed
from the experimental results are summarized in Table 2. The ten characteristics are
tabulated for each of the 48 games. The averages of these ten measures for each of the
eight experiments are also displayed. From these averages, it is possible to have some
idea about the effects the three factors on each of these measures. For instance, as one
moves from Run | to Run 2, the only input factor that changes is ‘demand pattern’
(from step-up to step-up-and-down). In this case, observe for example that the average
Max. Order measure changes from 143.3 to 146.7, a minor change, whereas the average
Max. Inventory changes from 98.3 to 167.5, clearly a more significant effect, at least at
this base level of the other factors, ignoring any interactions. (A simple t-test between
Run | and 2, ignoring other levels of the other two factors and possible interactions,
would yield the same conclusion). But a more complete and definitve conclusion about
the significance of each of the three factors on each of the ten output measures can be
obtained by a full analysis of variance (ANOVA), considering the effects of each factor
at all levels of other factors (and any possible interactions between them). A summary
table derived from full ANOVA (using SPSS software) is shown in Table 3. These
results are obtained from a full factorial ANOVA model involving seven effects (three
main effects, three 2-way interactions and one 3-way interaction term). Since we have a
total of 48 data, the degrees of freedom for residuals (errors) is 48-7-1= 40 and hence
the F statistic (~Mean Squared Explained by Regression/Mean Squared Error) for each
effect has (1, 40) degrees of freedom for numerator and denominator respectively. So, if
the F value computed for any effect is ‘large enough’, we reject the hypothesis that the
corresponding effect coefficient is zero, i.e. significant effect is discovered. Typical
significance levels used are alpha=0.01 (99% confidence)., alpha=0.05 (95%
confidence) or alpha=0.10 (90% confidence). In Table 3, we provide the F values
computed for each effect (for each output measure) and the ‘P value’ at which the F
value would be found significant. To conclude, if a P value is < the chosen alpha level,
we decide that the given effect has a significant effect on the selected output measure, at
(1-alpha)% confidence level. Although the ANOVA results comes from a full factorial
model, in Table 3, we show the main effects only for readability, because analyzing
each interaction term individually is beyond our research scope.
Ss 1: inventory 2: order
i 500.00 .
2: 250.00
1 0.00F EN N/A.
2 125.00
N 8 a 14:10 28 Ara 1998 Paz
FIGURE 4. Performance of the Player in Game 11 (Short Game with Orders Each
Period, Step Up and Down in Customer Demand, Exponential Delay)
(@ vinenoy 2: order
1 500.00
2 500.00
1 0.00.
2: 250.00
|
1 H
62.50 125.00 187.50 250.00
N a =a Graph 1: p9 (Subject 33) Time 15:18 28 Ara 1997 Paz
1
2: 0.00
FIGURE 5. Performance of the Player in Game 33 (Long Game with Orders Once
Every Five Periods, Step Up and Down in Customer Demand, Exponential Delay)
TABLE 2. Selected Measures of Game Performances
Min. | Max. | Range | mitial | Final |Duration) Inv.
Experiment Max. | Range | snven- | Invene (Oftnven-| Back | Back | Of | Oscil
Order /orOrders| Tver | Hyon JOrimwe™! order | order | Back | lation
Time _| Time | orders | Period
1 0 100 100 -125 50 175 4 38 34 N/A
2 0 100 100 -125 150 275 4 33 29 31
3 0 100 100 -175 150 325 4 18 14 29
4 0 200 200 -75 100 175 6 16 10 25
5 0 300 300 -225 100 325 5 43 38 N/A
6 20 60 40 -60 40 100 6 37 31 N/A
‘Avg Of Runt | 33 | 1433 | 1400 | -1308 | 983 | 2202 | 48 | 308 | 260 | 283
7 0 100 100 -60 115 175 10 21 IL 30
8 0 100 100 ~60 250 310 ) 13 8 21
9 0 60 60 -60 60 120 7 20 13 N/A
10 0 400 400, -80 180 260 6 21 15 39
ll 0 150 150 -100 250 350 6 14 8 28
12 40 70 30 -60 150 210 6 18 12 N/A
‘Avg OfRun2? | 67 | 1467 | 1400 | 700 | 1675 | 2375 | 67 | 178 | 112 | 295
13 0 100 100 300 525 5 34 29 43
14 0 220 220 225 450 x 35 30 45
15 0 150 150 160 235 14 25 ll N/A
16 0 60 60 190 215 Nié N/A N/A 25
17 0 100 100 250 375 3 38 33 N/A
18 0 300 300 180 440 5 18 13 27
Avg OfRun3 | 00 | 1550 | 1550 | i542 | 2175 | 3733 | 68 | 300 | 232 | 350
19 0 150 150 -25 150 175 N/A N/A NA N/A
20 0 150 150 -100 290 390 6 21 15 N/A
21 0 80 80 -145 80 225 ) 24 19 30
22 0 100 100 -40 350 390 12 18 4 32
23 i) 100 100 -250 210 460 6 27 21 45
24 0 200 200 -150 250 400 5 22 17 N/A
Avg of Runt | 00 | 1300 | 1300 | 1183 | 2217 | 3400 | 68 | 24 | 152 | 35.7
25 0 600 600 425 685 20 30 10 44
26 0 400 400, 200 550 25 85 60 N/A
27 0 300 300 -280 200 480 20 85 65 NIA
28 0 400 400 -300 200 500 25 70 45 63
29 0 560 560 -150 370 520 20 40 15 39
30 0 250 250 -20 425 445 35 45 10 73
Avg of Run5 | 00 | 483 | 4183 | 2267 | 3033 | 5300 | 242 | 592 | 342 | 548
31 40 375 335 -200 250 450 20 40 20 N/A
32 0 250 250 -100 200 300 25 55 30 N/A
33 0 400 400 -150 475 625 30 45 15 2B
34 50 250 200 -60 200 260 30 55 25 66
35 50 300 250 0 275 275 N/A N/A NIA N/A
36 100 250 150 -160 250 410 25 55 30 73
‘Avg Of Run6 | 400 | 3042 | 2042 | 1117 | 2750 | 3867 | 260 | 500 | 240 | 707
37 100 270 170 -100 500 600 20 40 20 85.
38 20 500. 480, -500 300 800 25 90 65 102
39 80 250 170 -40 450 490, N/A NA N/A N/A
40 50 250 200 -200 425 625 30 40 10 75
41 0 1000 1000 -1000 2000 3000 20 100 80 133
42 0 450 450, -375 420 795 35 60 25 52
Avg of Run? | 47 | 533 | 417 | 3092 | 0825 | 10517 | 260 | 060 | 400 | 894
43 0 250 250 -125 400 525 NA N/A N/A 75
44 0 450 450 -500 200 700 20 60 40 N/A
45 0 375 375 -750 200 950 20 70 50 N/A
46 20 300 280 -330 330 660 25 55 30 55
47 50 250 200 -180 300 480 30 45 15 114
48 0 300 300 -375 350 725 30 60 30 N/A
‘Avg Of Runé | 117 | 3208 | 3092 | 3767 | 2967 | 6733 | 250 | 380 | 330 | 813
TABLE 3. ANOVA Results (F values) on the Significance of the Effects of the Three Experimental Factors
Minimum | Maximum | RangeOf | Minimum | Maximum | Range Of Lae Final | puration of) EY
Factors Order Onder Orders Inventory | Inventory | Taventory ‘Time Time | Borders | period
B. 9.977 33.758 23.385 10.092 8.942 12.126 | 215,316 | 53.754 7.662 39.007
Length or | P Value
: 0.003 ) ° 0.003 0.005 0.001 0 0 0.009 °
Ordering |(v.=1v.=10}
accel significant | significant | significant | significant | significant | significant | significant | significant | significant | significant
Result | 4=0.05 a=0.05 a=0.05 a=0.05 a=0.05 a=0.05 a=0.05 a=0.05 a=0.05
= 0.013 0.137 0.103 4,383 6.782 0.346 1.048 0.637 5.626
oO P Value
Receiving 0.91 0.713 0.749 0.014 0.043 0.013 0.560 0.313 0.430 0.027
Dake [e1=hv2=40)
Result | insignificant | insignificant | insignificant Sieetiicant | sgniteant icant ficant ficant | “enificant
f significa sig Bare aeros | insignificant | insignificant | insignificant | “S™ OO!
Fo 0.084 3.003 2.667 1.125 1.596 0.325 4.539 4.228 0.144
Patten OF} Pvalue | 0.74 0.091 0.12 0.295 0.262 0.214 0.572 0.040 0.047 0.708
Customer |(v,=1Lv.=10)
Demand significant significant | significant
Resutt_ | insignificant | "8" | insignificant | insignificant insignificant | insignificant | “Ses same | insignificant
500.01
B® iv inventory 2: order
as
: 250.00]
2:
Z
b 7
1 -500.0 | \2 |
2: 0.001 } UU
0.00 25.00 50.00 75.00 100.00
N aa¥ Graph 1: p2 (Subject14) Time 12:42 23 Ara 1998 Car
FIGURE 6. Performance of the Player in Game 14 (Short Game with Orders Each
Period, Step Up in Customer Demand, Discrete Delay)
® 1: inventory 2: order
as 500.01
2 500.00
1
~4: 1
1: 0.00
2 250.00 VY
2
1 500.09 |
2: 0.00
0.00 62.50 125.00 187.50 250.00
N aaF Graph 1: p8 (Subject 43) Time 13:59 24 Ara 1997 Car
FIGURE 7. Performance of the Player in Game 43 (Long Game with Orders Once
Every Five Periods, Step Up and Down in Customer Demand, Discrete Delay)
Before we examine the ANOVA results of Table 3, note that the last four
mesures in Table 2 (Initial Backorder Time, Final Backorder Time, Duration of
Backorders nad Oscillation Period) have the symbol N/A in some of the cells. The
meaning is that the corresponding measure was simply undefined insome particular
game results. In a few instances, there are no backorders at all, so all three related
measures are marked N/A. A more problematic and frequent situation is when the
dynamics of inventory has no clearly identifiable, computable, constant period. This
arises when the inventory is non-oscillatory, or when it exhibits very complex, highly
noisy oscillations. In about two or three out of six runs in each cell we have this
situation, so it is quite frequent. Since ANOVA requires ‘balanced’ input data, these
N/A entries require some pre-treatment. Prior to ANOVA, SPSS uses a mixture of two
methods, depending on the suitability: ‘Estimating the missing data from neigboring
points’ or ‘discarding data from other cells so as to balance’. In any case, it should be
noted that the results about the last output measure, period of inventory oscillations,
should be taken with caution.
Although there are ten output measures in Table 2 and 3, some of these are
intermediate mesures used to compute related end-measures. Minimum and maximum
orders are measured to ultimately obtain a measure of ‘range of order fluctuation’ (or
amplitude) and the same is true for minimum and maximum inventories. Lastly, initial
backorder time and final backorder time are measured to compute a the ‘duration of
backorders.’ So to conserve space, we focus on four end-measures only and leave more
detailed examination to the interested reader.
3.1. Effect of Different Patterns of Customer Demand
ANOVA results in the bottom row block of Table 3 show the effects of
“demand pattern" (step-up only or step-up-and-down) on the output measures. As
illustrative game dynamics, see Figures 4, 5, 7 for step-up-and-down and Figures 6, 8B
and 10B for step-up demand. (For full results, compare pair wise the results of
experiments | and 2; 3 and 4; 5 and 6; 7 and 8 in Table 2. Also see Ozevin 1999), F-
values in Table 3 show that the demand pattern does not have a strong enough effect on
‘Range of Orders’ measure (although it does have some effect, since the significance is
just missed at 90%). Similarly, demand pattern has no significant effect on the
amplitude or period of inventory oscillations. Lastly, we observe that the demand
pattern does have a significant effect on the ‘duration of backorders’. This last finding
is interesting because the direction of the effect is that step-up-and-down demand,
compared to step-up only, causes the backlog durations to become shorter. So a
seemingly more complex demand pattern actually happens to compensate for the
players’ ordering weaknesses in terms of backlog durations.
3.2. Effect of Different Representations of Receiving Delays
Illustrative game dynamics with 1‘ order exponential delay and with discrete
delay are shown in Figures 4, 5 and Figures 6, 7 respectively. In these and many other
examples, with continuous exponential delay subjects are able to manage the
inventories in a relatively more stable way. It seems that when the goods ordered arrive
gradually over some period of time, it prevents the players from over-ordering or under-
ordering excessively. In contrast, discrete delay representation affects their performance
negatively by causing large fluctuations. In these experiments, subjects yield large
magnitude, long period oscillations in inventories and fail to bring these oscillations
11
under control in most cases. (For complete results, compare pair wise the results of
experiments 1 and 3; 2 and 4; 5 and 7; 6 and 8 in Table 2. Also see Ozevin 1999).
Subjects seem to have difficulty in accounting for the effects of sudden receiving. The
above results are consistent with research evidence on the effect of delays on dynamic
decision making performance, in system dynamics literature (Sterman 1989), as well as
in experimental psychology research (Brehmer 1989). The results are also consistent
with the mathematical stability conditions of discrete-delay dynamical systems, much
more difficult to obtain compared to continuous-delay systems. (See Driver 1977).
For a more statistical analysis, ANOVA results about the "type of receiving
delay effect" are given in the middle row block of Table 3. Observe that this factor has
very high significant effect on two critical output measures: Range of Inventory
fluctuations and its Period. This finding statistically confirms our qualitative assessment
above: that the type of receiving delay has a significant effect on the stability of
inventory fluctuations. (Also observe in Table 2 that the direction of the effect is that
discrete delay, compared to the continuous one, causes the range and period of
oscillations to be larger). In a nutshell, discrete delay representation makes the system
more oscillatory, less stable, and thus harder to manage for the subjects.
3.3. Effect of the Length of Decision Intervals
As illustrative game dynamics with ‘order every time step’ and ‘order every
five time steps’ experiments; see Figures 4, 6 and Figures 5, 7 respectively. (For full
results, compare pair wise the results of experiments 1 and 5; 2 and 6; 3 and 7; 4 and 8
in Table 2. Also see Ozevin 1999). These comparisons reveal that all output measures
change significantly, when the ordering interval is changed from ‘every time step’ to
every five time steps’. ANOVA results given in the first row block of Table 3 are also
consistent with this observation: Effects on all measures are highly significant. There
may be two different sources of this high level and comprehensive significance: First,
when the decision period is made longer (noting that the receiving delay is also made
longer to be consistent), since some time-constants of the system are larger, time-related
output measures (like Period of Oscillations, Backorder Times and Backorder
Duration) all naturally become larger in value — a natural technical result of dynamics
of the system. Secondly —and more interestingly- amplitude measures (like Range of
Orders and Range of Inventory) also become statistically larger. This is more
behavioral/decision-related result: When decisions can be made less frequently, orders
must be larger in magnitude, feedback is less frequent, controlling of the inventory
becomes harder and so inventory fluctuations are larger in magnitude as well. So again
in sum, ‘order every five time steps’’ makes the system more oscillatory, less stable,
thus harder to manage for the subjects.
4, Testing of the Alternative Decision Formulations
Section 3 above, completes the relatively more quantitative/statistical research
objective of the paper. In light of the above results, the second objective is to evaluate
the adequacy of standard decision rules typically used in dynamic stock management
models and to seek improvement formulations. To this end, the performance patterns of
subjects will be (qualitatively) compared with the dynamic patterns obtained using
different simulated ordering formulations: As a preliminary step, the patterns of
ordering behavior of subjects are observed to fall in three basic classes: i- smooth,
continuous (oscillatory or non-oscillatory) damping orders (for example Figure 4), ii-
12
alternating large discrete and then zero orders, like a high frequency signal (for example
Figure 6), iii- long periods of constant orders punctuated by a few sudden large ones
(for example Figure 10-B). Using these three types of observed ordering patterns, the
common linear "Anchoring and Adjustment” rule, several “nonlinear” adjustment rules
and some standard discrete inventory control rules (such as (s, Q)) found in the
inventory management literature will be evaluated.
4.1. Linear Anchoring and Adjustment Rule
Linear Anchoring and Adjustment Rule is frequently used to model decision-
making behavior in System Dynamics models (Sterman, 2000, 1987). In making such
decisions, one starts from an initial point, called the anchor, and then makes some
adjustments to come up with the final decision. In the context of inventory management,
a plausible anchor point for order decisions is the expected customer demand. (If the
inventory manager can order only once every five periods, the anchor should be the
total of expected customer demand for five periods between subsequent decisions.)
When there are discrepancies between desired and actual inventory levels and/or
between desired and actual supply line, adjustments are made so as to bring the
inventory and the supply line back to desired levels. Thus, the order equation based on
linear Anchoring and Adjustment heuristic is formulated as:
O.= E+ a*(Le-I) + B*(SL:- SL) (4.1)
When orders can be given once every five periods, then the anchor of the rule is
modified as follows (the adjustment terms being as before):
O= 5*E¢ (Ir -1,) + B*(SLi - SL) (4.2)
where E; represents expected customer demand, I represents the desired inventory, I;
the inventory, SL,’ the desired supply line and SL; the supply line. a and B are the
adjustment fractions.
In real life, ‘safety stocks’ are determined by balancing the inventory holding
and backordering costs. Although, an optimum inventory level minimizing these costs
may be found mathematically, more often safety stocks are set approximately. The
desired inventory I," is thus modeled as proportional to customer demand to allow
adjustments in safety stocks when changes in customer demand occur.
I, =k*E, (4.3)
To maintain a receiving rate consistent with receiving delay t and customer
demand, SL, is formulated as a function of t and the expected customer demand E,.
SL = 1*E; (4.4)
The linear anchor and adjust rule can mimic the subjects’ performances
adequately in experiments where they tend to place smooth and continuously damping
orders (See Figure 8 as an example and Ozevin 1999 for more). However, in certain
game conditions, most subjects tend to order non-continuously (for example Figure 6
and 10B), especially when the receiving delay representation is discrete and/or the order
interval is five. Such order patterns typically fall in class ii (alternating large discrete
and then zero orders, like a high frequency signal) or class iii (long periods of constant
orders punctuated by a few sudden large ones) described above. The linear "Anchoring
and Adjustment Rule" can not yield such discrete-looking intermittent or occasional-
large-then-constant orders. Non-linear rules may be able to represent better the
subjects’ performances in such situations. Different approaches will be discussed below
in two separate sections: Non-linear adjustment rules and standard inventory control
tules.
13
B inventory 2: order
1
500.004
2 250.00]
L o.oo} Re a cae me
2 125.00] ¥ TY
+} =a
2
aos an ea
1 -500.00}
2 0.00. .
0.00 25.00 50.00 75.00 100.00
GEX (A) Graph 2 (IAdjt=2, SLAdjt=50) Time 00:19 01 Oca 1994 Cum
” 1: inventory 2: order
1 500.0
2: 250.00)
rh
a ae
0.0} ans SS
125.0
1:
2:
PW nn,
1: -500.0 ii yw
2 0.00;
(8) 25.00 50.00 75.00 100.00
N 8 =Fal Graph 1: p5 (Subject 3) Time 12:45 25 Ara 1998 Cum
FIGURE 8. Comparison of (A) the Anchoring and Adjustment rule (a = 0.5, B= 0.02),
with (B) the Performance of the Player in Game 3 (Short Game with Orders Each
Period, Step Up in Customer Demand, Exponential Delay)
14
4.2. Rules with Nonlinear Adjustments
The linear adjustment rule is based on ‘adjustments’ to orders proportional to
the discrepancy between the desired and observed stock levels. Orders are placed
regularly, in proportion to this discrepancy. However, some players do not place such
smooth orders. They completely cease ordering when the inventory seems to be ‘around
a satisfactory’ level and place rather large orders as the discrepancy between the desired
and actual inventory becomes larger. The resulting ordering behavior is in class iii (long
periods of constant orders punctuated by a few sudden large ones) described above. In
this section, three alternative nonlinear decision rules will be developed and tested to
address this particular class of nonlinear ordering behavior.
4.2.1, Cubic Adjustment Rules
Similar to the Linear Anchoring and Adjustment rule, Cubic Adjustment rules
also start with expected customer demand as an anchor point, but the adjustments are
formulated as non-linear. One or both of the adjustment terms may be cubic in
discrepancies (in inventory and/or in supply line. Alternative order equations can be
mathematically expressed as:
OF E+ F(t - 1)? (45)
if only the inventory adjustment is taken into account, and:
O.=Er+ a¥*(Ir - 1)? + B*(SL:- SL: (4.6)
O.= Et oF(lr - 1) + B*(SLr- SL)? (4.7)
where one of the adjustments is made cubic and the other is linear and finally,
O. =E, + a*(I" - 1)° + B*(SL- SLY? (4.8)
where both of the adjustments are formulated as cubic. In the equations above, E;
represents the expected customer demand; I, and SL," the desired inventory and supply
line levels; I; and SL; the actual inventory and supply line and. a and are the fraction
of the discrepancy corrected by the decision-maker at each period. The internal
consistency of the rules can be shown mathematically (See Ozevin 1999). When orders
can be given once every five periods, the adjustments terms are as above; however, the
anchor term is increased to five times the expected customer demand.
The general stability properties of the cubic adjustment rules are similar to the
well established linear adjustment rule. In other words, firstly the supply line must be
taken into account (i.e. adjustment fraction B must be non-zero and preferably equal to
a for optimum stability) and secondly, the larger the values of a and B the less stable
the system tends to be. For example, Figure 9 compares two behaviors of the cubic rule
with non-zero and zero B values. Observe that the behavior becomes quite unstable
when the supply line adjustment term is zero. Although the behavior in Figure 9 (A)
illustrates a very stable case, we should note that choosing stable values for the
adjustment fractions is not easy with Cubic Adjustment rules (whereas the linear rule
guarantees stability for any a=f < 1). In other words, the performance of the cubic rule
is too sensitive to the adjustment parameter values. In particular, when orders are given
once every five periods, the cubic rule most often fails to generate stable dynamics. On
the other hand, the primary advantage of the cubic rule is that it is possible to generate
ordering patterns that can mimic the nonlinear ordering behavior of subjects
characterized by long periods of constant orders punctuated by a few sudden large ones
(described as class iii, above). Figure 10 provides comparison for such a case.
Although cubic adjustment formulations can potentially yield such nonlinear order
patterns with discrete delays as well as continuous ones, the range in which they can
15
yield stable dynamics is too narrow to be useful. (See Ozevin 1999). Hence,
motivation for other nonlinear rules is to be discussed in the following sections.
YD 2: inventory
2: order
1: 500.004
2: 250.00
{ ~~ Ls
r? rh
0.00,
2 125.00] “y
2 tio ——
‘ies pe ar
L -500.00
0.00
0.00 25.00 50.00 75.00 100.00
82 (A) SLCe: pl (SLAdjt=1/1500, |Adjime 22:43 06 Feb 2004 Fri
1: inventory 2: order
1 500.004
2 250.00
al
i
[1
1: 0.00, 2
Py 125.00] | }
1 -500.00
2 0.00: 2 2
() 25.00 50.00 75.00 100.00
Ne af LICPSLCe: pl ( Adjt=1/1500,SLAdjime 22:45 06 Feb 2004 Fri
the
FIGURE 9. Two Different Performances of the “Cubic Supply Line and Cubic
Inventory Adjustment rule” with two different parameters: (A) (with a=1/500,
B=1/1500) and (B) (with a=1/500, B=0).
16
BP 2 inventory 2: order
1
: 500.00)
2: 250.00
2
1 \_| L 3.
r1- ia | SS
1 0.00,
2 125.00
s/ eee reese
1 500.00)
0.00.
0.0 A 25.00 50.00 75.00 100.00
BF (A) ICe: p2 (SLAdjt=1, IAdjt=500)Time 02:27 20 Oca 1994 Per
é] 1: inventory 2: order
1: 500.0
2: 250.00)
1: 0.00)" ee
2: 125.00
Y
2
ae
£5000
: : 25.00 50.00 75.00 100.00}
Nese h:pl(Subject1) Time 12:32 04 Eyl 1986 Per
FIGURE 10. Comparison of (A) Performance of “Linear Supply Line and Cubic
Inventory Adjustment rule” (a=1/500, B=1) with (B) the Performance of the Player in
Game 1 (Short Game with Orders Each Period, Step Up in Customer Demand,
Exponential Delay)
17
4.2.2. Variable Adjustment Fraction Rule
Analogous to the Anchoring and Adjustment Rule and the Cubic Adjustment
Rules, we define a ‘Variable Adjustment Fraction Rule’ that anchors at expectations
about customer demand. However, the adjustments are increased sharply (nonlinearly)
when the discrepancy in inventory increases. The simplest version of the order equation
of the rule can be mathematically expressed as
OK E+ a*(h-1) (4.9)
ion
adjustment
-3 2 1 0) a 2 3
normalized inventory discrepancy
FIGURE 11. Graphical Adjustment Fraction Function
where the variable fraction a is a function of the discrepancy in inventory, normalized
by the desired inventory. The shape of the function yields increased adjustments when
the discrepancy in inventory is increased. Normalized inventory discrepancy 6 is
defined as
O= a (4.10)
According to the function in Figure 11, the rule is not mathematically unbiased
in the ideal known demand case; there will be some small, deliberate steady state
discrepancy between the inventory and its desired level. But this may well be a
"realistic" bias in order to be able to obtain a non-linear ordering behavior similar to
some subjects. The rule performs quite realistically in the "noisy" demand case, where
the steady state bias is negligible anyway and may be irrelevant in real life. The more
general (and more stable) version of the Variable Adjustment Fraction Rule will be of
the form E; + a*(I;-I) + B*(SL:- SL), where B is defined by a function exactly
equivalent to the one in Figure 11, except that the input would be ‘normalized supply
line discrepancy.’ Alternatively, the supply line adjustment may be linear. (We omit this
discussion further in this article to conserve space). In any case, inclusion of the supply
line term will increase the stability of the system.
18
Variable Adjustment Fraction rule typically generates long periods of constant
orders punctuated by a few sudden large ones. Therefore they may be used to represent
subjects’ behavior when orders are characterized by this type of nonlinearity, where
linear adjustment rules would fail (See Figure 12 and Ozevin 1999 for more
illustrations).
WD vinenory order
y 500.004
330.00
a a a oe
F 00 99
z ‘a, + 1
doo 2500 soa 75:00 10000
hase (A) carssivesneyemincin re ves) oases
.) 1: inventory 2: order
1 500.01
2 250.00
ee oe
L 0.00)"
2 125.00
—2-~—_~2
‘ -500.00] ANE |
25.00 50.00 75.00 100.00
Née@
FIGURE 12. Comparison of (A) Variable Adjustment Fraction rule with (B) the
Performance of the Player in Game 10 (Short Game with Orders Each Period, Step Up
and Down in Customer Demand, Exponential Delay)
(B) 1: p9 (Subject 10) Time 11:35 28 Ara 1998 Paz
19
4.2.3. Nonlinear Expectation Adjustment Rule
The order equation of another nonlinear rule that we call ‘expectation
adjustment rule’ can be mathematically expressed by
O, = a*E, (4.11)
where the variable adjustment coefficient a is a function of discrepancy in inventory
normalized by desired inventory and E; represents the expected customer demand. o is
equal to one when the inventory is at the desired level, since adjustments are not
necessary when the system is in equilibrium (See Figure 13). The shape of the a
function causes increasing upward adjustments in orders when the inventory level is
below the desired level and it causes reductions in orders when the inventory is above
its desired level.
adjustment coeffici
—
oO
ot
P/N FH ©
L
2 1 0 1 2 3 4
normalized inventory discrepancy
FIGURE 13. Adjustment Coefficient Function for the Non-linear Expectations
Adjustment Rule
Like the previous Variable Adjustment rule, Nonlinear Expectations Rule can
yield infrequent large orders. Therefore, it may provide an alternative to the linear
anchoring and adjustment rule when subjects’ behaviors exhibit such patterns (See
Figure 14 for example).
Finally, this section yields another general result: the well-documented
oscillatory dynamic behavior of the inventory when supply line is underestimated is true
not only for the linear anchor-and-adjust rule but also for the non-linear rules seen
above. Furthermore, in non-linear rules, stability is achieved for a rather narrow range
of parameter/function values.
20
f*} 1: inventory 2: order
ty
2
500.09)
250.00}
Pa
1 o.ogf ™\ Mgt. Lt
2: 125.09}
2 iN a BAN
L 500.09 / 2
2: 0.0
9,00 25.00 50.00 75.00 100.0
a =F) (A) inventory and orders exponential : irtite 00:08 02 Feb 2004 Mon
Ss 1: inventory 2: order
L 500.01
2 250.00
1 0.00] + a
2 125.00
2)
1-~—2-
2
25.00 50.00 75.00 100.00
1: p11 (Subject 11) Time 14:10 28 Ara 1998 Paz
FIGURE 14. Comparison of (A) Nonlinear Expectation Rule with (B) the Performance
of the Player in Game 11 (Short Game with Orders Each Period, Step Up and Down in
Customer Demand, Exponential Delay).
21
5. Standard Inventory C ontrol Rules
The third type of non-linear behavior, alternating large discrete and then zero
orders of players and the resulting zigzagging inventory patterns, suggests that discrete
inventory control rules used in inventory management may be suitable. The following
four policies are most frequently used in inventory management literature:
Order Point, Order Quantity (s, Q) Rule;
Order Point, Order Up to Level (s, S) Rule;
Review Period, Order Up to Level (R, S) Rule;
(R, s, S) Rule.
Two fundamental questions to be answered by any inventory control system
are “how many” and “when” (or “how often”) to order. “Order-Point” systems
determine how many to order, whereas “Periodic-Review” systems determine how often
to order as well (Silver and Peterson, 1985), (Tersine, 1994). When subjects can order
every time unit, they are free to order at any time they desire. Therefore, order-point
systems, rather than periodic review systems are more appropriate to represent the
ordering behavior in these situations. In contrast, when subjects can order only, say,
once every five periods, periodic-review systems with five as review period may be
more appropriate as decision rules. These inventory control rules assume that time flow
and changes are discrete. Therefore, these rules will be tested only with discrete delays.
As will be seen, these rules are non-linear in the sense that they consist of piecewise,
discontinuous functions.
5.1. Order Point-Order Quantity (s, Q) Rule:
Order Point-Order Quantity (s, Q) rule can be mathematically expressed as
O.=Q, ifEh<s
0, otherwise (5.1)
where EI, represents the effective inventory and s the “order point”. Effective inventory
and order point are calculated as follows:
El =1+ SLi (5.2)
DAVGSL, + DMINI; (5.3)
I, and SL; represent goods in inventory and in supply line respectively. DAVGSL, refers
to the desired average supply line and DMINI, to desired minimum inventory. To be
consistent with the previous continuous adjustment rules, desired average supply line
and desired minimum inventory can be defined as
DAVGSL, = t*E; (5.4)
DMINI,= E;, + SS (5.5)
in terms of receiving delay t, expected demand E, and safety stocks SS. Order quantity
Q and safety stock SS are fixed arbitrarily as constants in this research. Desired
minimum inventory DMINI; is defined as the sum of a constant safety stock and
demand expectation. With such a definition, desired minimum inventory can be adapted
to variations in customer demand. The performance of this rule with deterministic
demand is seen in Figure 15. Orders exhibit alternating zeros and Q’s and inventory
22
zigzags around a constant level. Observe that the (s, Q) rule can not prevent the
inventory from falling below the desired minimum inventory, even when no noise
exists, primarily because the order quantity Q is constant. This particular rule is
therefore not suitable for our purpose (i.e. for comparative evaluation against the
continuous stock adjustment rules). (s, Q) rule will not be further evaluated; it is
defined only as a preliminary for the other more realistic rules to follow.
s 1: inventory 2: DMINI 3: order
x 200.00
a: 250.00]
_—_—"
3 0.00]
ae 125.00
| Ei
3}
2 200.00 i
3 0.00
0.00 25.00 50.00 75.00 100.00
Nee# Graph 2: p2 (SQ rule2) Time 03:17 20 Oca 1994 Per
FIGURE 15. Performance of Order Point-Order Quantity Rule (Q=4*Djo, DAVGSL; =
4*D,, DMINI = 0) in Short Game with Orders Each Time Unit, Known Step Up in
Customer Demand, Discrete Delay.
5.2. Order Point-Order Up to Level (s, S) Rule
Order Point-Order Up to Level (s, S) rule can be mathematically expressed as
O,=S-Eh, if El<s
0, otherwise (5.6)
where EI, represents the effective inventory, s the order point and S the upper level of
inventory. Effective inventory, the order point and the upper level S of inventory can be
defined as follows:
EI, =1,+ SL; 65.7)
s = DAVGSL, + DMINI (5.8)
S=s+Q (5.9)
I, and SL; represent goods in inventory and in supply line respectively. DAVGSL; refers
to the desired average supply line and DMINI, to desired minimum inventory. Desired
average supply line and desired minimum inventory are defined as before:
DAVGSL, = t*E; (5.10)
DMINI,= E:+ SS (5.11)
23
in terms of receiving delay t, expected demand E, and safety stocks SS. Order size Q
and safety stock SS are initially set as constants. But note that the actual order quantity
O, (eq. 5.6.) is a variable in this rule. Desired minimum inventory DMINI, is defined as
the sum of expectations and safety stocks. As such, desired minimum inventory DMINI,
may be adapted to variations in customer demand. This rule can be shown to be
unbiased and mathematically consistent in the deterministic case. (See Ozevin 1999).
The inventory reaches equilibrium at the desired minimum inventory level, but in the
“noisy” case it may move up and down around the desired minimum, due to differences
between the "expected" and "actual" customer demands (Figure 16).
The (s, S) rule takes fully into account the supply line in the sense that the
decision is based on the effective inventory EI = I + SL. Thus, in accordance with the
fundamental result for the linear anchor and adjust rule, the resulting inventory
dynamics is non-oscillatory. The zigzagging behavior of the inventory seen in Figure 16
is caused by the time-discrete, piecewise ordering rule yielding discrete, alternating zero
and non-zero orders. (There is also some minor wavelike dynamics caused simply by
the autocorrelated random demand). To explore this point further, a modified version
the (s, S) rule is run by redefining EI = I + k*SL, where k is the supply line inclusion
coefficient. Normally, for full inclusion of supply line k is one. To demonstrate partial
inclusion of supply line, the dynamics of (s, S) model is illustrated with k=0.70 in
Figure 17. Observe that the inventory now does exhibit an oscillatory pattern. So the
hypothesis: ‘players that generate oscillatory inventories ignore/underestimate the
supply line’ is compelling in general, whether the ordering heuristic is linear, non-linear
or piecewise/discontinuous.
s 1: inventory 2: DMINI 3: order
y 200.005
2
3: 250.00
1
2 7 2
y 0.00 y
3 125.00
1 F}
L
Hl 200.00] ii
3 0.004
0.00 25.00 50.00 75.00 100.00
NeeF Graph 2: p6 (SS rule4NE) ‘Time 04:26 20 Oca 1994 Per
FIGURE 16. Performance of Order Point-Order Up to Level (s, S) Rule (with
DAVGSL, = 4*E,, DMINI = E, EI, = I, + SL,) in Short Game with Orders Each Time
Unit, Step Up in Customer Demand, Discrete Delay.
v) 1: inventory 2: DMINI 3: order
4 200.005
2
3 250.00
Fp ae
. — _— ia
: ee 2 2 2
; 0.00]
eH 125.00
3}
| 200.00 ;
3: 0.00 3. 3
0.00 25.00 50.00 75.00 100.00
Néeae $SRule: p3 (C=0.7) Time 00:18 02 Feb 2004 Mon
FIGURE 17. Oscillatory Performance of modified Order Point-Order Up to Level (s, S)
rule (DAVGSL; = 4*E;, DMINI = E, and El, = I, +k* SL; where k=0.70) in Short Game
with Orders Each Time Unit, Step Up in Customer Demand, Discrete Delay.
5.3. Review Period, Order Up to Level (R, S) Rule
Order Point-Order Up to Level rule can be mathematically expressed as:
O. = S-El, ift=R*k
0 otherwise (5.12)
where EI; represents the effective inventory, t the time, S the upper level of inventory. k
is an integer and R is the review period. Effective inventory, the order point and the
upper level S of inventory are defined as follows
EI, =1,+ SL; (5.13)
S =DAVGSL, + DMINI, + R*E; (5.14)
I, and SL; represent goods in inventory and in supply line respectively. Review period R
is five. DAVGSL, refers to the desired average supply line and DMINI; to desired
minimum inventory. Desired average supply line is defined as
DAVGSL, = t*E; (5.15)
in terms of receiving delay t, expected demand E;. Desired minimum inventory DMINI;
corresponds to the safety stock. DMINI, is arbitrarily fixed as constant.
This rule can be shown to be unbiased and mathematically consistent in the
deterministic case. (See Ozevin 1999). The inventory reaches an equilibrium point
consistent with the desired minimum inventory level, but in the “noisy” case it may fall
below the desired minimum due to noise effects. (Figures 18 and 19 provide two
illustrations). The basic behaviors are again alternating large discrete and then zero
orders and zigzagging inventory patterns.
25
F) 1: inventory 2: DMINI 3: order
y 500.004
2
3:
1000.00
3
A
2 -500.00
3: 0.00
0.00 62.50 125.00 187.50 250.00
INGT=fa Graph 5: p5 (RS rule5NE) Time 00:52 01 Oca 1994 Cum
FIGURE 18. Performance of Review Period-Order Up to Level (R,S) rule (with
DAVGSL, = 10*E,, DMINI = 0) in Long Game with Orders Every Five Time Units,
Step Up in Customer Demand, Discrete Delay.
F3 1: inventory 2: DMINI 3: order
yj 500.004
2.
3 1000.00
1: \
1
3
0.00 62.50 125.00 187.50 250.00
N 6eF Graph 5: p6(RS rule6NE) Time 00:51 01 Oca 1994 Cum
FIGURE 19. Performance of Review Period-Order Up to Level (R,S) rule (with
DAVGSL, = 10*E,, DMINI = 0) in Long Game with Orders Every Five Time Units,
Step Up and Down in Customer Demand, Discrete Delay.
26
5.4. (R,s, S) Rule:
Finally, (R, s, S) rule can be mathematically expressed as
O,=S-El, ift=k*R and EI <s
0 otherwise (5.16)
where S represents the upper level of inventory, El, the effective inventory, R the
review period, t the time and s the order point. Review period R is five. Effective
inventory, the order point, the upper level S of inventory and safety stock SS are:
El, =I+ SL (5.17)
SS=R*E,+DMINI, (5.18)
s = DAVGSL, + SS; (5.19)
S=s+R*E, (5.20)
I, and SL; represent goods in inventory and in supply line respectively. DAVGSL, refers
to the desired average supply line, DMINI, to desired minimum inventory, SS to safety
stock, E; to expected demand. Desired average supply line is defined as:
DAVGSL, = t*E; (5.21)
in terms of receiving delay t, expected demand E,. DMINI, is determined as a constant.
This rule can be shown to be unbiased and mathematically consistent in the deterministic
case (See Ozevin 1999). Due to the difference between the expected and actual customer
demand in the noisy case, (R, s, S) rule may not result in ordering each time the effective
inventory falls to or below the order point; orders are sometimes delayed until the
following period. The inventory reaches an equilibrium consistent with the desired
minimum inventory level, but in the "noisy" case it may occasionally fall below the
desired minimum. (See Figure 20). The fundamental patterns of behavior are similar to
those seen with the previous discrete inventory rules.
i: inventory 2: DMINI 3: order
4 500.00
2
3 1000.00
1
Wy, 1
1
‘ 0.00, y
3 500.00
3
1 500,09 Ml H AL 4 LIL LJ
0.00 62.50 125.00 187.50 250.00
ING T=fa Graph 7:p4(RSS rule4NE) Time 01:46 01 Oca 1994 Cum
FIGURE 20. Performance of (R, s, S) Rule (DAVGSL, = 10*E, , DMINI = 0) in Long
Game with Orders Once Every Five Time Units, Step Up in Customer Demand,
Discrete Delay.
27
5.5. Comparison of the Standard Inventory Rules with Game Results
Order point-Order Quantity (s, Q) rule is not a plausible decision rule
formulation when demand is not constant, as discussed above. Order Point-Order Up to
Level (s, S) rule on the other hand may provide an adequate representation of subjects’
performance in some cases where the continuous Anchoring and Adjustment Rule is
inadequate. One such case is depicted in Figures 21 and 22. Observe that the ordering
behavior of subjects is discontinuous, consisting of alternating large discrete and then
zero orders, well represented by the (s, S) orders of Figure 21. The zigzagging
inventory patterns of Figure 21 and 22 are also consistent. The linear or nonlinear
continuous adjustment rules on the other hand can not produce such behavior patterns,
especially if orders are allowed in each time step.
The ordering patterns generated by the Review Period, Order Up to Level (R,
S) rule on the other hand are generally quite similar to the ones produced by the
anchoring and adjustment rules when the value of R is matched with minimum ordering
interval (5 days in our experiments). Therefore, (R, S) rule does not provide novel
behavior patterns that anchoring and adjustment rules fail to represent.
But (R, s, S) rule can represent subjects’ decisions where anchoring and
adjustment rules fail, especially in cases where intervals between subjects’ non-zero
orders are not constant. One such comparison is depicted in Figures 23 and 24. Note
that continuous anchor and adjust rules, when applied under ‘orders are given every five
time steps’ condition, would fail to generate variable time intervals in between non-zero
orders.
28
1 inventory 2: DMINI 3: order
y 200.004
2
3 250.00
1
\ ;
Li
ae
x 0.00] :
3 125.00
3
1 3
2 200.00
3: 0.004
0.00 25.00 50.00 75.00 100.00
N 6 =a Graph 2: p7 (SS ruleSNE) Time 04:25 20 Oca 1994 Per
FIGURE 21. Performance of Order Point-Order Up to Level (s, S) Rule (DAVGSL, =
4*E,, DMINI = E,) in Short Game with Orders Each Time Unit, Step Up and Down in
Customer Demand, Discrete Delay.
I 1: inventory 2: order
1: 500.0
2: 250.00]
1:
2: 125.00]
0.00/72 NWN YANW
a" |
Ss
L -500.00]
2: 0.00 2 2. 2
0.00 25.00 50.00 75.00 100.00
N aaFf Graph 1: p8 (Subject20) Time 10:23 24 Ara 1998 Per
FIGURE 22. Performance of the Player in Game 20 (Short Game with Orders Each
Time Unit, Step Up and Down in Customer Demand, Discrete Delay).
29
s 1: inventory 2: DMINI 3: order
4] 800.004
2
3: 500.00
onl MUNDY Nw iyeP MAN
62.50 125.00 187.50 250.00
Graph 7: p5 (RSS rule5NE) Time 23:23 01 Feb 2004 Sun
FIGURE 23. Performance of (R, s, S) Rule (DAVGSL; = 10*E,;, DMINI = 0) in Long
Game with Orders Once Every Five Time Units, Step Up and Down in Customer
Demand, Discrete Delay.
F*) 1: inventory 2: order
L 800.0
2 500.0
be
L 0.0
2 250.0 bi \ D
1
2
lL -800.0 |
2 0.00-——2 2 2
0.00 62.50 125.00 187.50 250.00
NéeF Graph 1: p13 (Subject 45) Time 12:57 09 Eyl 1986 Sal
FIGURE 24. Performance of the Player in Game 45 (Long Game with Orders Once
Every Five Time Units, Step Up and Down in Customer Demand, Discrete Delay).
30
6. Conclusion
This paper has two different research objectives: The first one is to analyze the
effects of selected experimental factors on the performances of subjects (players) in a
stock management simulation game. To this end, the generic stock management
problem is chosen as the interactive gaming platform. Gaming experiments are designed
to test the effects of three factors on decision making behavior: different patterns of
customer demand, minimum possible order decision (‘review’) interval and finally the
type of the receiving delay. ANOVA results of these 3-factor, 2-level experiments show
which factors have significant effects on ten different measures of behavior (such as
max-min range of orders, inventory amplitudes, periods of oscillations and backlog
durations).
The second research objective is to evaluate the adequacy of standard decision
rules typically used in dynamic stock management models and to seek improvement
formulations. The performances of subjects are compared against some selected
ordering heuristics (formulations). First, a classification of the patterns of ordering
behavior of subjects reveals three basic types: i- smooth oscillatory (or non-oscillatory)
damping orders, ii- alternating large and zero orders, like a high frequency discrete
signal, iii- long periods of constant orders punctuated by a few sudden large ones.
Comparing these behaviors with the stand-alone simulation results, we observe that the
common linear "Anchoring and Adjustment Rule." can mimic well the first (i) type of
behavior, but can not, due to its linear nature, replicate the next two types. Several
alternative non-linear rules are formulated and tested against subjects’ behaviors. Some
standard discrete inventory control rules (such as (s, Q)) common in the inventory
management literature are also formulated and tested. The non-linear adjustment rules
are found to be more representative of subjects’ decisions in cases where subjects’
ordering patterns fall in class (ili) above. The standard discrete inventory rules are more
representative of the subjects’ ordering behavior in cases where decision patterns
consist of alternating discrete large and zero orders. Another finding is the fact that the
well-documented oscillatory dynamic behavior of the inventory is a quite general result,
not just an artifact of the linear anchor and adjust rule. When the supply line is ignored
or underestimated, large inventory oscillations result, not just with the linear anchor-
and-adjust rule but also with the non-linear rules, as well as the standard inventory
management rules. Furthermore, depending on parameter values, nonlinear ordering
rules are more prone to instability -even if the supply line is taken into account.
More research is needed to formulate and test other non-linear formulations.
There is also need to test these rules in more complex and realistic game environments
(such as more stocks, delays and multi-player supply chains). Currently, we are
examining the effects of information delays in the order decisions (in addition to supply
line delays). Initial results indicate that the oscillations become more unstable and the
basic anchor and adjust heuristic needs to be modified properly in order to reduce the
instability. Finally, a long term research question is: if individuals tend to order
intermittently and discrete delays further exacerbate the situation, are mostly continuous
system dynamics decision structures invalid? Or is it true that even if individuals decide
in discrete and intermittent fashion, the macro decision making behavior of the
aggregate system (that we are typically interested in) can be/should be modeled
continuously? There is a fundamental research question here that should bridge the
micro behavior of actors (using perhaps agent-based modeling) to the macro behavior of
the aggregate system.
31
References
Barlas, Y. and A. Aksogan, “Product Diversification and Quick Response
Order Strategies in Supply Chain Management,” Proceedings of the International
System Dynamics Conference, 1997
Barlas Y. and M.G. Ozevin. 2001. "Testing The Decision Rules Used In Stock
Management Models” Proceedings of the 19th International System Dynamics
Conference, Atlanta, 2001
Brehmer, B., “Feedback Delays and Control in Complex Dynamic Systems,”
in Milling, P. and E. Zahn (eds.), Computer Based Management of Complex Systems,
Berlin: Springer-Verlag, pp.189-196, 1989.
Driver, R.D. Ordinary and Delay Differential Equations, Springer-Verlag,
Berlin, 1977.
Forrester, J.W., Industrial Dynamics. Pegasus Communications,,
Massachusetts, 1961.
Lee, H.L., Padmanabhan V. and Wang S., “The Bullwhip Effect in Supply
Chains”, Sloan Management Review 1997.
Ozevin, M. G.,Testing The Decision Rules Frequently Used In System
Dynamics Models”, M.S. Thesis, Bogazigi University, 1999.
Sterman, J. D., “Testing Behavioral Simulation Models by Direct
Experiment,” Management Science, Vol. 33, No.12, pp. 1572-1592, December 1987.
Sterman, J. D., “Modeling Managerial Behavior: Misperceptions of Feedback
in a Dynamic Decision Making Experiment,” Management Science, Vol. 35, No. 3, pp.
321-339, March 1989.
Sterman, J.D. Business Dynamics. Systems Thinking and Modeling for a
Complex World. McGraw-Hill, U.S.A., 2000.
Tersine, R. J., Principles of Inventory and Materials Management, Prentice-
Hall, Englewood Cliffs, New Jersey, 1994.
Tvede, Lars. Business Cycles: The Business Cycle Problem from John Law to
Chaos Theory. Harwood Academic Publishers, Switzerland, 1996.
Silver, E. A. and R. Peterson, Decision Systems for Inventory Management
and Production Planning, John Wiley and Sons, New York, 1985.
Yasarcan H. and Barlas Y. , “A General Stock Control Formulation For stock
Management Problems Involving Delays And Secondary Stocks,” Proceedings of the
International System Dynamics Conference,Palermo, 2002.
32
Appendix A: Interactive Inventory Management Game Interface
LSSTELLA® 5.01 te) ee) ad SSI EE IIS Ke)/) PEI ES)
File Edt Map Bun Help
IMGc5s.STM
es
raph
ZoomTo
Graph
MStart| Epouum Penceresi |[-RsTELLAe 5.11
33
Appendix B: Game Instruction Sheet
Inventory Management Game (Short Version)
1. Objective
In this game, as an inventory manager, you will control the inventory level of
a certain good such that your company must not backlog too many orders from
your customers. You can achieve this by keeping a large safety stock. However,
large safety stocks result in high inventory costs. Therefore, you should keep
your inventory level as low as possible while trying not to backlog.
supplyline inventory
a?) © omnia
order receiving delivery
2. How the inventory is controlled?
You will control your inventory by ordering new goods. You can order new
goods once every day. While ordering new goods you should consider the
following three variables: inventory, demand and supply line. Inventory is the
quantity of goods you have in hand. Demand is the quantity of goods requested
by within a time unit (a day). If there are no enough goods in your inventory at
any time, you will take the order as a backlog and supply the goods later. In this
case, your inventory will be negative until you receive enough goods from your
suppliers. Supply line corresponds to the goods you have ordered previously but
you have not received yet. Remember that there are time delays between your
placing of orders and receiving them.
3. Remarks
e There is no cost associated with ordering. Therefore you may order as
frequently as every time period.
e By inspecting the graph of inventory and your orders over time, you may
have some idea about the supply line delay and the future demand.
e You can enter your order either by typing it or by sliding the input device.
After entering your order, press the play button.
e The game will last 100 days. At the end, please save your game under C:
\STELLAS\ as yourname.stm
34
Appendix C: Model Equations (Illustrated with linear anchor and adjust rule)
Demand Sector
demand = step_input +noise
noise = SMTH1(whitenoise,2,0)
step_input = 20+STEP(20,4)-STEP(20,19)
whitenoise = NORMAL(0,0.15*step_input,67779)
Goal Sector
desired_inventory = inventory_coefficient*step_input
desired_onorder = step_input *receiving delay
inventory_coefficient = 2
Order Sector
adjtimel =5
adjtime2 = 3
decision_rule = demand+inventory_adjustmenttsupplyline_adjustment
inventory_adjustment = inventory_discrepancy/adjtime2
supplyline_adjustment = supplyline_discrepancy/adjtimel
Production Line Sector
inventory(t) = inventory(t - dt) + (receiving - delivery) * dt
INIT inventory = inventory_coefficient*step_input
INFLOWS:
receiving = supplyline/receiving delay
OUTFLOWS:
delivery = demand
supplyline(t) = supplyline(t - dt) + (order - receiving) * dt
INIT supplyline = 80
INFLOWS:
order = decision_rule
OUTFLOWS:
receiving = supplyline/receiving_delay
inventory_discrepancy = desired_inventory-inventory
receiving delay =4
supplyline_discrepancy = desired_onorder-supplyline
SEE Back to the Top
CC BY-NC-SA 4.0