Available
Available online
online at
at www.sciencedirect.com
www.sciencedirect.com
ScienceDirect
ScienceDirect
Available
online atonline
www.sciencedirect.com
Available
at www.sciencedirect.com
ScienceDirect
ScienceDirect
Procedia
Procedia CIRP
CIRP 00
00 (2019)
(2019) 000–000
000–000
www.elsevier.com/locate/procedia
www.elsevier.com/locate/procedia
Procedia CIRP
00 (2017)
000–000
Procedia
CIRP 88
(2020) 580–583
www.elsevier.com/locate/procedia
13th
13th CIRP
CIRP Conference
Conference on
on Intelligent
Intelligent Computation
Computation in
in Manufacturing
Manufacturing Engineering,
Engineering, CIRP
CIRP ICME
ICME ‘19
‘19
Restaurants
management
on
demand
forecasting
28thstore
CIRP Design
Conference,based
May 2018,
France
Restaurants
store
management
based
on Nantes,
demand
forecasting
a,
aa
bb
c
Tanizaki
,, Takeshi
,, Takeshi
Takenaka
A newTakashi
methodology
analyzeHoshino
the functional
and physical
of
Takashi
Tanizakia,*,
*,toTomohiro
Tomohiro
Hoshino
Takeshi Shimmura
Shimmura
Takeshiarchitecture
Takenakac
Graduate
School
of
University,
739-2116,
Graduatefor
Schoolan
of Kindai
Kindai
University, 1
1 Takaya-Umenobe,
Takaya-Umenobe,
Higashi-Hiroshima
739-2116, Japan
Japan
existing products
assembly
orientedHigashi-Hiroshima
product family
identification
Ritsumeikan University, 1-1-1 Nogi-Higashi, Kusatsu 525-8577, Japan
a
a
c
c
b
b
Ritsumeikan University, 1-1-1 Nogi-Higashi, Kusatsu 525-8577, Japan
National
National Institute
Institute of
of Advanced
Advanced Industrial
Industrial Science
Science and
and Technology,
Technology, 6-2-3
6-2-3 Kashiwanoha,
Kashiwanoha, Kashiwa,
Kashiwa, Chiba,
Chiba, 277-0882,
277-0882, Japan
Japan
Paul Stief *, Jean-Yves Dantan, Alain Etienne, Ali Siadat
*
author. Tel.:
+81-82-434-7484;
fax: +81-82-434-7890.
E-mail address:
tanizaki@hiro.kindai.ac.jp
* Corresponding
Corresponding
Tel.:Supérieure
+81-82-434-7484;
+81-82-434-7890.
address:
tanizaki@hiro.kindai.ac.jp
Écoleauthor.
Nationale
d’Arts etfax:
Métiers,
Arts et MétiersE-mail
ParisTech,
LCFC
EA 4495, 4 Rue Augustin Fresnel, Metz 57078, France
* Corresponding author. Tel.: +33 3 87 37 54 30; E-mail address: paul.stief@ensam.eu
Abstract
Abstract
In
In this
this paper,
paper, restaurants
restaurants store
store management
management based
based on
on demand
demand forecasting
forecasting is
is proposed.
proposed. The
The restaurant
restaurant service
service industry
industry has
has low
low productivity
productivity due
due
Abstract
to
to the
the simultaneity
simultaneity of
of service
service goods.
goods. In
In order
order to
to solve
solve such
such problems,
problems, we
we are
are researching
researching how
how to
to manage
manage restaurant
restaurant stores
stores such
such as
as employee
employee
placement,
placement, food
food material
material ordering,
ordering, etc.,
etc., based
based on
on highly
highly accurate
accurate demand
demand forecasting
forecasting by
by machine
machine learning
learning with
with internal
internal data
data such
such as
as POS
POS data
data
Inand
today’s
business
environment,
the trend
towards
moreand
product
variety
andpaper,
customization
is the
unbroken.
Dueresults
to this of
development,
the need
of
external
data
exiting
in
ubiquitous
such
as
weather
events.
In
this
we
discuss
forecasting
customer
order
and external data exiting in ubiquitous such as weather and events. In this paper, we discuss the forecasting results of customer order quantity
quantity
agile
and reconfigurable
production
systems
emerged
to cope with
various
products
and
productforfamilies.
Tochain
design
and optimize production
and
shop
inventory
order
quantity
of
draft
beer
using
forecasting
method
with
machine
learning
restaurant
R.
and shop inventory order quantity of draft beer using forecasting method with machine learning for restaurant chain R.
systems
well
as to choose
the by
optimal
product
© 2020
2019as
The
Authors.
Published
Elsevier
B.V. matches, product analysis methods are needed. Indeed, most of the known methods aim to
2019
The
Authors.
Published
by Elsevier
Elsevier
B.V.
©
The
Authors.
Published
by
B.V.
analyze
a
product
or
one
product
family
on
the
physical
level.
Different
families, however,
may differ
largely ininterms
of the number and
Peer-review
under
responsibility
of
the
scientific
committee
of
the
13thproduct
CIRP Conference
Conference
on Intelligent
Intelligent
Computation
Manufacturing
This
is an open
access
article under
BY-NC-ND
license
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review
under
responsibility
of the CC
scientific
committee
of
the
13th
CIRP
on
Computation
in Manufacturing
nature
of components.
This
fact impedes
anscientific
efficient committee
comparisonofand
choice
of appropriate
product
family Computation
combinations inforManufacturing
the production
Engineering.
Peer
review
under
the
responsibility
of
the
the
13th
CIRP
Conference
on
Intelligent
Engineering.
system.
A new17-19
methodology
is proposed
to analyze
existing products in view of their functional and physical architecture. The aim is to cluster
Engineering,
July 2019,
Gulf of Naples,
Italy.
these
products
in new
assemblyMachine
orientedlearning;
productRandom
familiesforest
for the
optimization
existing assembly
lines
and the creation of future reconfigurable
Keywords:
Demand
forecasting;
regression;
Service
Restaurant
management
Keywords:
Demand
forecasting;
Machine
learning;
Random
forest
regression;
Serviceofengineerring;
engineerring;
Restaurant
management
assembly systems. Based on Datum Flow Chain, the physical structure of the products is analyzed. Functional subassemblies are identified, and
a functional analysis is performed. Moreover, a hybrid functional and physical architecture graph (HyFPAG) is the output which depicts the
similarity between product families by providing design support to both, production system planners and product designers. An illustrative
1.
such
data
and
in
example
of a nail-clipper is used to explain the proposed methodology. An internal
industrial data
case study
product
of steeringdata
columns
of
1. Introduction
Introduction
internal
data
suchonas
astwoPOS
POS
datafamilies
and external
external
data
in the
the
thyssenkrupp Presta France is then carried out to give a first industrial evaluation
of
the
proposed
approach.
ubiquitous
environment
such
as
weather,
events,
etc.
in
order
ubiquitous environment such as weather, events, etc. in order to
to
© 2017
The
Authors.
Published
by
Elsevier
B.V.in today's developed
With
the
globalization
of
the
economy,
improve
With the globalization of the economy, in today's developed
improve the
the accuracy
accuracy of
of forecasting.
forecasting. We
We have
have verified
verified the
the
Peer-review
under responsibility
of the
scientific
committee oftotheGDP
28th CIRP
Design Conference
countries including
Japan, the
ratio
of manufacturing
effectiveness
of the2018.
forecasting method with the statistical
countries including Japan, the ratio of manufacturing to GDP
and
has
declining.
It
and employment
employment
has been
been
declining.
It is
is the
the service
service industry
industry
Keywords:
Assembly; Design
method;
Family identification
that
absorbs
employment
in
place
of
manufacturing
that absorbs employment in place of manufacturing industry
industry in
in
developed
developed countries
countries [1].
[1]. In
In Japan,
Japan, the
the service
service industry
industry accounts
accounts
for
for about
about 70%
70% of
of GDP
GDP and
and employment.
employment. On
On the
the other
other hand,
hand, the
the
productivity
of
the
service
industry
is
lower
than
that
1.productivity
Introduction
of the service industry is lower than that of
of the
the
manufacturing
manufacturing industry.
industry. Therefore,
Therefore, improving
improving the
the productivity
productivity
of
the
industry
an
in
to the
fast is
in the
domain the
of
ofDue
the service
service
industry
isdevelopment
an important
important issue
issue
in revitalizing
revitalizing
the
entire
Japanese
economy.
In
order
to
solve
such
problems,
it
communication
and an ongoing
trend
digitization
entire Japanese economy.
In order to
solveofsuch
problems,and
it is
is
important
the
of
in
digitalization,
manufacturing
enterprises
areimprovement
facing important
important to
to change
change
the method
method
of business
business
improvement
in the
the
service
from
“experience
and
challenges
in today’s
market
environments:
continuingto
service industry
industry
from
“experience
and aintuition”
intuition”
to
“engineering
method”.
From
the
above
background,
we
are
tendency
towards
reduction
of product
development
times
“engineering
method”.
From
the above
background,
weand
are
researching
the
improvement
of
productivity
of
shortened
product
In addition,
is an increasing
researching
the lifecycles.
improvement
of the
the there
productivity
of the
the
restaurant
using
the
engineering
method.
In
particular,
we
demand
of customization,
being atmethod.
the same
time in a global
restaurant
using the engineering
In particular,
we are
are
researching
how
to
management
improving
competition
all over
the world.by
trend,
researching with
how competitors
to advance
advance store
store
management
byThis
improving
employees'
work
arrangement
and
food
materials
which
is inducing
the development
from materials
macro toordering
micro
employees'
work arrangement
and food
ordering
based
accurate
number
for
markets,
in forecasting
diminished of
due of
to customers
augmenting
based on
onresults
accurate
forecasting
oflotthe
thesizes
number
of
customers
for
face-to-face
service
industries,
especially
restaurants.
As
part
product
varieties
(high-volume
low-volume
production)
[1].of
face-to-face
service
industries, to
especially
restaurants.
As part
of
the
research,
we
are
forecasting
using
To
with this
variety
as well asmethods
to be able
to
thecope
research,
we augmenting
are researching
researching
forecasting
methods
using
identify possible optimization potentials in the existing
2212-8271
2019
Published
Elsevier
B.V.
production
is important
toby
have
a precise
2212-8271 ©
©system,
2019 The
TheitAuthors.
Authors.
Published
by
Elsevier
B.V. knowledge
effectiveness of the forecasting method with the statistical
method
method and
and machine
machine learning
learning using
using the
the above
above data
data in
in the
the visitor
visitor
number
forecasting
for
restaurant
chain
R
[2].
In
this
number forecasting for restaurant chain R [2]. In this paper,
paper, we
we
discuss
discuss the
the forecasting
forecasting results
results of
of customer
customer order
order quantity
quantity and
and
shop
shop inventory
inventory order
order quantity
quantity of
of draft
draft beer
beer (hereafter
(hereafter beer)
beer)
using
forecasting
method
with
machine
learning
for
ofusing
the forecasting
product range
and characteristics
manufactured
and/or
method
with machine learning
for restaurant
restaurant
chain
assembled
chain R.
R. in this system. In this context, the main challenge in
modelling and analysis is now not only to cope with single
2.
Method
products,
a limited
product range or existing product families,
2. Forecasting
Forecasting
Method
but also to be able to analyze and to compare products to define
this
used
as
new In
product
families. machine
It can be learning
observedis
classical
existing
In
this research,
research,
machine
learning
isthat
used
as forecasting
forecasting
method.
Machine
learning
is
a
method
to
find
regular
patterns
product
arelearning
regrouped
function
clients
or features.
method.families
Machine
is ainmethod
toof
find
regular
patterns
inherent
in
data
by
learning
data
iteratively.
By
applying
new
However,
assembly
families are
to find.
inherent in
data by oriented
learningproduct
data iteratively.
By hardly
applying
new
data
to
the
learning
results,
it
is
possible
to
forecast
the
future
Ontothe
level,
differ
mainly
two
data
theproduct
learningfamily
results,
it is products
possible to
forecast
theinfuture
according
to
Although
methods
have
main
characteristics:
(i) the
number various
of components
(ii)been
the
according
to the
the pattern.
pattern.
Although
various
methodsand
have
been
developed
for
learning,
we
forest
type
of components
(e.g. mechanical,
electronical).
developed
for machine
machine
learning, electrical,
we use
use random
random
forest
regression
this
Classicalin
regression
inmethodologies
this research.
research. considering mainly single products
or solitary, already existing product families analyze the
product structure on a physical level (components level) which
causes difficulties regarding an efficient definition and
comparison of different product families. Addressing this
Peer-review
of
the
of
2212-8271
2020responsibility
The Authors.
bycommittee
Elsevier B.V.
Peer-review©under
under
responsibility
of Published
the scientific
scientific
committee
of the
the 13th
13th CIRP
CIRP Conference
Conference on
on Intelligent
Intelligent Computation
Computation in
in Manufacturing
Manufacturing Engineering.
Engineering.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer review
under
the
responsibility
scientific
2212-8271
© 2017
The
Authors.
Publishedofbythe
Elsevier
B.V.committee of the 13th CIRP Conference on Intelligent Computation in Manufacturing
Engineering,
17-19
July 2019,ofGulf
of Naples,
Italy. of the 28th CIRP Design Conference 2018.
Peer-review
under
responsibility
the scientific
committee
10.1016/j.procir.2020.05.101
Takashi Tanizaki et al. / Procedia CIRP 88 (2020) 580–583
T. Tanizaki et al./ Procedia CIRP 00 (2019) 000–000
2.1. Random Forest
Random forest is an ensemble learning method that
constructs a forest using multiple decision trees and integrates
learning results for each decision tree [3]. Random forest is an
ensemble learning method that constructs a forest using
multiple decision trees and performs majority decision on the
result of learning for each decision tree. In order to prevent
extreme bias in learning of each decision tree, learning data
used in each decision tree is extracted with randomness. As a
result, overfitting is prevented and high generalization
performance is obtained.
581
classified into the same category. In addition, beer Z is not sold
at the other three stores except A.
Table 1. Explanatory variable.
2.2. Random Forest Regression[ 3 ]
There are two types of methods in the random forest:
“classification” and “regression” [3]. The difference between
the two methods is that “classification” is used if data can be
divided into classes, and “regression” is used to forecast
continuous data such as time series. In this research, we use
random forest regression because we forecast time series data.
3. Forecast Customer Order Quantity and Shop
Inventory Order Quantity
3.1. Target Data
Customer order quantity and inventory order quantity for
beer (beer X, beer Y, beer Z) in 4 stores of A, B, C, D are
forecasted using visitor record data, customer order data, and
inventory order data of restaurant chain R. A machine learning
model using random forest regression is constructed in Python.
This model is learned using actual data of '15/5/1 to '17/4/30.
Using this learning result, we forecast the customer order
quantity and inventory order quantity for beer from '17/5/1 to
'18/ 4/30, and compare with actual results. Table 1 shows the
explanatory variables used for forecasting. The forecasting
ratio α is calculated using Eqs. (1) and (2).
pi : Actual value on i-th day
ei : Forecasting value on i-th day
N : Forecast period
αi : Forecasting ratio on i-th day
𝑝𝑝𝑝𝑝𝑖𝑖𝑖𝑖 − |𝑝𝑝𝑝𝑝𝑖𝑖𝑖𝑖 − 𝑒𝑒𝑒𝑒𝑖𝑖𝑖𝑖 |
𝑝𝑝𝑝𝑝𝑖𝑖𝑖𝑖
∑𝑁𝑁𝑁𝑁
𝑖𝑖𝑖𝑖=1 𝛼𝛼𝛼𝛼𝑖𝑖𝑖𝑖
α=
𝑁𝑁𝑁𝑁
𝛼𝛼𝛼𝛼𝑖𝑖𝑖𝑖 =
(1)
(2)
The inventory ordering unit for beer in restaurant chain R is
a barrel (19L/barrel). On the other hand, Table 2 shows
volumes of beer W, beer X, beer Y, beer Z. Therefore, in the
inventory order quantity forecasting, the customer order
quantity is converted to the inventory order quantity. Beer W
has started handling at each store from '17/9/1. Since there is
no beer W order before '17/8/31, beer W is in the same category
as beer X. Before the sale of beer W, beer X was a standard
volume product in restaurant chain R. Since beer W is a product
with a slight increase in volume, it is considered to be similar
as a customer's order motive, so these two products were
Table 2. Volumes of beer.
Category
Volume(mL)
W
420
X
360
Y
220
Z
135
3.2. Forecasting Results of Customer Order Quantity and
Inventory Order Quantity
The customer order quantity of the three items and the
inventory order quantity are forecasted in the following two
cases.
(1) Forecasted quantities using 2016 data (case1).
Takashi Tanizaki et al. / Procedia CIRP 88 (2020) 580–583
T. Tanizaki et al. / Procedia CIRP 00 (2019) 000–000
582
(2) Forecasted quantities using data between 2015 and 2016
(case2).
Table 3 shows the forecasting results. The forecasting ratios
for customer order quantity and inventory order quantity are
about 30% to 70% and are not high. There are some stores
where the forecasting ratio of the learning model using data for
two years is higher than that of the learning model using data
for one year, but some stores are low.
Table 3. Forecasting results.
Customer order
Category
Store
X
A
54.9
50.5
B
36.4
57.0
quantity
Y
Case1(%) Case2(%)
C
58.4
65.2
D
67.8
65.0
A
41.6
45.9
B
55.8
35.0
C
52.0
55.3
D
46.6
49.0
Z
A
31.0
22.0
Inventory oder quantity
A
71.0
70.6
B
53.4
52.3
C
45.1
53.5
D
49.8
48.6
Fig. 1 shows the forecasting result for customer order
quantity of category X at store A using the learning model of
Case 1, and Fig. 2 shows that of Case 2. The red line is the
forecasting value, and the blue line is the actual value. There is
a difference between the actual value and the forecasting value
in both Fig.1 and Fig.2. The forecasting value captures the
trend of increase and decrease of actual value. However, it
fluctuates around the average value, and large fluctuation of the
actual value cannot be forecasted. The forecasting value of
Case 1 which is a learning result using one-year actual data is
slightly less different from the actual value than the forecasting
value of Case 2 which is a learning result using two-year actual
data. The forecasting result of Case 2 with more data is worse.
Fig. 1. Forecasting result for customer order quantity of Case 1 at store A.
Table 4 shows the characteristics of actual value for
customer order quantity in 2015-2017. The total of customer
order quantity in 2015 is the largest in three years except
category Y store C and category Y store D. Among the 5 cases
in which total value for customer order quantity in the past 2
years have been decreasing, there are 3 cases where the
forecasting ratio is higher using one-year actual data. In this
case, it is considered that machine learning using data of the
one-year actual data is expected to have a higher forecasting
ratio, but there are cases where it does not. Among the 4 cases
without the above tendency, there are 3 cases where the
forecasting ratio is higher using two-year actual data. In this
case, it is considered that machine learning using data of the
two-year actual data is expected to have a higher forecasting
ratio, but there are cases where it does not. Since the forecasting
is based on machine learning results for two years, there is a
possibility that the relationship between the data transition and
the forecasting ratio is not captured because the number of data
is small. In the future, we will increase the number of years of
learning and investigate changes in the forecasting ratio.
Table 4. Characteristic of actual value for customer order quantity.
Total of customer order quantity
Category
Store
X
A
15895
12765
12015
Case1
B
15975
14229
13929
Case2
C
26332
25674
24184
Case2
D
23524
20524
21651
Case1
A
1641
880
1044
Case2
Case1
Y
Z
2015
2016
2017
Tendency Higher ratio
B
931
806
750
C
2455
2016
3105
Case2
D
2244
1219
2565
Case2
A
770
431
418
Case1
Fig. 3 shows the forecasting result for inventory order
quantity at store A using the learning model of Case 1, and Fig.
4 shows that of Case 2. The red line is the forecasting value,
and the blue line is the actual value. The actual result is an
intermittent order system in which there are days when the
order is executed and days when the order is not executed. That
is, beer to be sold on multiple days is ordered at one time. As a
result, the order quantity on the day on which the order is
executed is higher than forecasting. On the other hand,
forecasting result is almost always ordering the same amount.
The actual ordering method is not a good ordering method
because a large warehouse capacity is required. Therefore, we
examined which of the actual ordering method and the
forecasting ordering method matched the customer order
quantity.
Fig. 2. Forecasting result for customer order quantity of Case 2 at store A.
Fig. 3. Forecasting result for inventory order quantity of Case 1 at store A.
Takashi Tanizaki et al. / Procedia CIRP 88 (2020) 580–583
T. Tanizaki et al./ Procedia CIRP 00 (2019) 000–000
583
Table 5. Fitting ratio of category X.
Store
Case1(%) Case2(%) Case3(%)
A
53.7
57.6
31.1
B
56.7
55.1
23.7
C
59.6
56.5
49.8
D
51.5
47.8
46.6
5. Conclusion
Fig. 4. Forecasting result for inventory order quantity of Case 2 at store A.
4. Comparison of inventory order quantity and customer
order quantity of next day
Based on the previous discussion, the matching ratio of
inventory order quantity customer order quantity was
examined. The beer arrival lead time at Restaurant Chain R is
one day. If the inventory order is ordered to match the customer
order quantity of the next day, the inventory quantity can be
reduced. Therefore, the actual inventory order quantity and
forecasting result for inventory order quantity were compared
with the customer order quantity for the next day. The fitting
ratio β is calculated using Eqs. (3) and (4).
qi : Actual customer order value on i-th day
fi : Forecasting value for inventory order (or actual inventory
order) on i-th day
N : Verification period
βi : Fitting ratio on i-th day
𝑞𝑞𝑞𝑞 −|𝑞𝑞𝑞𝑞𝑖𝑖𝑖𝑖 −𝑓𝑓𝑓𝑓𝑖𝑖𝑖𝑖−1 |
𝛽𝛽𝛽𝛽𝑖𝑖𝑖𝑖 = 𝑖𝑖𝑖𝑖
𝛽𝛽𝛽𝛽 =
𝑞𝑞𝑞𝑞𝑖𝑖𝑖𝑖
∑𝑁𝑁𝑁𝑁
𝑖𝑖𝑖𝑖=2 𝛽𝛽𝛽𝛽𝑖𝑖𝑖𝑖
𝑁𝑁𝑁𝑁 − 1
(3)
(4)
We compared the actual inventory order and forecasting
results of inventory order by random forest regression in the
same way as Chapter 3.
(1) Forecasted quantities using 2016 data (case1).
(2) Forecasted quantities using data between 2015 and 2016
(case2).
(3) Actual quantities of 2017 (case3)
Table 5 shows fitting ratio of category X of each case.
Machine learning models have a higher matching ratio to
customer orders than actual methods. It is considered that
the inventory amount can be reduced by placing an
inventory ordering using a machine learning model
because the difference from the customer order amount
on the next day by the machine learning model is smaller.
However, the fitting ratio is about 60% and is not high. In
the future, we will work on upgrading the machine
learning model to improve the fitting ratio.
The random forest regression, which is one of machine
learning, was used to forecast the customer order quantity and
the inventory order quantity of beer at each store of the
restaurant chain R. We use internal data such as POS data and
external data in the ubiquitous environment such as weather,
events, etc. in order to improve the accuracy of forecasting.
Based on the results, we compared the actual inventory
ordering method and the inventory ordering method according
to this research.
(1)Forecasting results of customer ordering amount
The forecasting ratio was 22% to 68% and the accuracy was
not high. In the comparison of forecasting ratio using learning
data for one year and learning data for two years, results
differed depending on the store, and systematic findings were
not obtained. As the influence of single-year data may be
strong, we will increase the number of years of learning data
and verify.
(2)Forecasting results of inventory ordering amount
The forecasting ratio was 45% to 71% and the accuracy was
not high. Since the actual inventory ordering method is not
good, we decided to investigate fitting ratio to the customer
order amount of the next day.
(3)Comparison of inventory order method
The fitting ratio for the customer order amount on the next
day is 48% to 60% in the forecasting method and 24% to 50%
in the actual method, and the fitting ratio was higher in the
forecasting method.
In the future, we will increase the number of years of
learning data and investigate the effects on forecasting ratio and
fitting ratio, and prepare for practical use.
Acknowledgements
This study is supported by JSPS KAKENHI (16H02909).
References
[1] Mitsutaka M. Policy Challenges to Promote Service Innovation. The
journal of the Institute of Electronics, Information and Communication;
96(8):638-642(In Japanese)
[2] Takashi T, Tomohiro H, Takeshi S, Takeshi T. Demand forecasting in
restaurants using machine learning and statistical analysis; Procedia
CIRP;79:679-683
[3] Sebastian R. Python Machine Learning (Japanese Edition); Impress
Corp:86-87