Available online at www.sciencedirect.

com
Available online at www.sciencedirect.com
Available ScienceDirect
online atonline
www.sciencedirect.com
ScienceDirect
Available at www.sciencedirect.com

ScienceDirect
Procedia CIRP 00 (2018) 000–000
ScienceDirect
Procedia CIRP 00 (2018) 000–000 www.elsevier.com/locate/procedia
www.elsevier.com/locate/procedia
Procedia CIRP 00 (2017)
Procedia 000–000
CIRP 79 (2019) 679–683
www.elsevier.com/locate/procedia

12th
12thCIRP
CIRPConference
Conferenceon
onIntelligent
IntelligentComputation
ComputationininManufacturing
ManufacturingEngineering,
Engineering,18-20
CIRP July
ICME2018,
'18
12th CIRP Conference on Intelligent Computation in Manufacturing
Gulf of Naples, Italy Engineering, CIRP ICME '18
Demand
28th CIRPforecasting in restaurants
Design Conference, using
May 2018, Nantes, France
Demand forecasting in restaurants using
machine learning and statistical analysis
A new methodologymachine to analyze
learningthe
andfunctional
statisticaland physical architecture of
analysis
existing products
Takashi for
Tanizaki a,
a,
an assembly
*,Tomohiro oriented
Hoshino a
a
,Takeshi product
Shimmurabfamily
b
,Takeshi identification
Takenakac
c
Takashi Tanizaki *,Tomohiro Hoshino ,Takeshi Shimmura ,Takeshi Takenaka
Graduate Schoole of Kindai University, 1 Takaya-Umenobe, Higashi-Hiroshima 739-2116, Japan
a

Paul Stief
Graduate
a
Schoole *, Jean-Yves
Ritsumeikan
b
University,
of Kindai University, Dantan,
1-1-1 Alain
Noji-Higashi,
1 Takaya-Umenobe, Etienne,
Kusatsu Ali
525-8577, Japan
Higashi-Hiroshima Siadat
739-2116, Japan
Advanced Industrial
c b Science
Ritsumeikan and Technology,
University, 2-3-26 Aomi
1-1-1 Noji-Higashi, Koto-Ku,
Kusatsu Tokyo Japan
525-8577, 135-0064, Japan
École Nationale Supérieure
c d’Arts
Advanced et Métiers,
Industrial Arts and
Science et Métiers ParisTech,
Technology, LCFC
2-3-26 AomiEA 4495, 4Tokyo
Koto-Ku, Rue Augustin Fresnel,
135-0064, Japan Metz 57078, France
* Corresponding author. Tel.: +81-82-434-7484; fax: +81-82-434-7890. E-mail address: tanizaki@hiro.kindai.ac.jp
* Corresponding author. Tel.: +81-82-434-7484; fax: +81-82-434-7890. E-mail address: tanizaki@hiro.kindai.ac.jp
* Corresponding author. Tel.: +33 3 87 37 54 30; E-mail address: paul.stief@ensam.eu

Abstract
Abstract
Abstract
In this paper, demand forecasting in restaurants using machine learning is proposed. Many researches have been proposed on demand forecasting
technology
In this paper,using POSforecasting
demand data. However, in orderusing
in restaurants to make demand
machine forecasts
learning at a realMany
is proposed. store,researches
it is necessary to establish
have been proposeda on
store-specific demand
demand forecasting
In today’s business
forecasting
technology model in
using POSenvironment,
consideration theoftrend
data. However, towards
various
in orderfactorsmore
to makesuchproduct
as the variety
demand store andatcustomization
location,
forecasts store, itisisevents,
the weather,
a real unbroken.
etc. Due
necessary to this we
Therefore,
to establishdevelopment,
constructedthe
a store-specific need of
a demand
forecasting
agile model inthatconsideration
and reconfigurable functionally
production of combines
systems the
variousemergedabove
factors mentioned
to cope
such datalocation,
withstore
as the varioususing machine
products andlearning.
product
the weather, In this
etc. paper,
families.
events, the demand
To design
Therefore, forecasting
andconstructed
we optimize model
production
a demand
using
systems machine
as well
forecasting learning
as that
model and the
to choose theverification
functionally optimal
combines result
the of
product the model
matches,
above usinganalysis
product
mentioned real using
data storemethods
data is discussed.
machine arelearning.
needed. In Indeed, most the
this paper, of the known
demand methods aim
forecasting to
model
© 2018
using
analyze aThe
machine Authors.
learning
product or onePublished
and
product by Elsevier
the verification
family B.V.
result
on the of thelevel.
physical model using real
Different store data
product is discussed.
families, however, may differ largely in terms of the number and
Peer-review
©
© 2018
nature
2019ofThe under
components.
The responsibility
Authors.
Authors. Published
This fact by
Published of Elsevier
the scientific
impedes
by ElsevieranB.V. committee
efficient
B.V. of the 12th
comparison and CIRP
choiceConference
of appropriateon Intelligent Computation
product family in Manufacturing
combinations for the production
Engineering.
Peer-review
system. A new
Peer-review under responsibility
methodology
under of the
is proposed
responsibility scientific
to analyze
of the scientific committee
existing
committee of the
ofproducts12th CIRP
in view
the 12th CIRP Conference on
of their functional
Conference Intelligent Computation
andComputation
on Intelligent in Manufacturing
physical architecture. The aimEngineering.
in Manufacturing is to cluster
Engineering.
these products in new assembly oriented product families for the optimization of existing assembly lines and the creation of future reconfigurable
Keywords:systems.
assembly Demand Based
forecasting, Machine
on Datum learning,
Flow Chain,Statistical analysis,
the physical Serviceof
structure engineerring,
the productsRestaurant management
is analyzed. Functional subassemblies are identified, and
a Keywords:
functionalDemand forecasting,
analysis Machine
is performed. learning, aStatistical
Moreover, analysis, Service
hybrid functional engineerring,
and physical Restaurantgraph
architecture management
(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
example of a nail-clipper is used to explain the proposed methodology. An industrial case study on two product families of steering columns of
1. Introduction
thyssenkrupp Presta France is then carried out to give a first industrial evaluation researching demand approach.
of the proposed forecasting methods using internal data
1. Introduction
© 2017 The Authors. Published by Elsevier B.V. researching
such as POSdemanddata forecasting
and external methods
data using
in theinternal data
ubiquitous
The service industry is an important industry accounting for such as
environment
Peer-review under responsibility of the scientific committee of the 28th CIRP Design Conference 2018. POS suchdataas and
weather, external
events, data
etc. in
in the
order ubiquitous
to improve
The 70%
about serviceofindustry
Japan'sis anGDP.important
However,industry accounting
since the laborfor environment
the accuracy of such as weather,
demand events,Inetc.
forecasting. thisinpaper,
orderwe
to describe
improve
about 70%
Keywords:
productivity of the
Assembly;
of Japan's
Design GDP.
method;
service However,
Family
industry lower since
identification
is than the oflabor
that the the approach
accuracy of to demand forecasting.
forecasting In thisonpaper,
method based we describe
machine learning
productivity
manufacturing ofindustry,
the service industry is lower
its improvement than that policy
is an important of the the
and approach
statisticaltoanalysis.
forecasting method based on machine learning
manufacturing industry, itsinimprovement
issue of Japan. Especially the labor-intensiveis an important policy
service industry and statistical analysis.
issue of
labor Japan. Especially
productivity is low in the labor-intensive
because service goodsservice industry
are consumed 2. Forecasting method
1. Introduction of the product range and characteristics manufactured and/or
labor
at theproductivity
same time isaslow theybecause service goods
are provided. Among are the
consumed
labor- 2. Forecasting method
assembled in this system. In this context, the main challenge in
at the same
intensive time industries,
service as they arerestaurant
provided. industry,
Among the whichlabor-
is In this research, the number of customers is forecasted using
Due to the fast development in the domain of modelling and analysis is now not only to cope with single
intensive
integrated service
production industries,
and sales restaurant
industry,industry,
has improvedwhich its is In thislearning
machine research,and
the statistical
number ofanalysis
customers is forecasted
method using
with internal
communication and an ongoing trend of digitization and products, a limited product range or existing product families,
integrated production
inventory possibilitiesand by sales industry,service
separating has improved
production its machine
data learning data
and external and in
statistical analysisenvironment.
the ubiquitous method withBayesian
internal
digitalization, manufacturing enterprises are facing important but also to be able to analyze and to compare products to define
inventory possibilities
functions from by separating
sales by introducing a centralservice production
kitchen. On the data
LinearandRegression,
external dataBoosted
in the ubiquitous
Decision environment. Bayesian
Tree Regression, and
challenges in today’s market environments: a continuing new product families. It can be observed that classical existing
functions
other hand,fromin sales by introducing
the face-to-face a central
service kitchen.
industry like On the
dinner Linear Regression,
Decision Boosted are
Forest Regression Decision Treemachine
used for Regression, and
learning,
tendency towards reduction of product development times and product families are regrouped in function of clients or features.
other hand, separating
restaurants, in the face-to-face service industry
service production functionslikefromdinner
sales Decision Forest Regression
Stepwise method is used for are used for
statistical machine
analysis learning,
method. We
shortened product lifecycles. In addition, there is an increasing However, assembly oriented product families are hardly to find.
restaurants,
is separating
not applicable servicethe
because production
quality functions
of servicefrom willsales
be Stepwise
used Azuremethod
Machineis used for statistical
Learning as a machineanalysis method.
learning We
tool and
demand
is of customization,
not applicable becausetobeing
the atquality
the same time in awillglobal On the product family level,asproducts differ mainly in two
compromised. In order solve such ofproblems,
service we arebe used
SPSSAzure Machine
as a statistical Learning
analysis tool. a machine learning tool and
competition
compromised. withIncompetitors all over the problems,
world. Thiswe trend, main
SPSScharacteristics: (i) the number
tool. of components and (ii) the
researching how to orderadvance to store
solve such
management by improving are as a statistical analysis
which is inducingto the
researching development from macro to micro type of components (e.g. mechanical, electrical, electronical).
employees' how advance
work arrangement store
andmanagement
food materialsby improving
ordering
markets,
employees'results in diminished
work forecasting
arrangementofand lot sizes due to
food materials augmenting
ordering Classical methodologies considering mainly single products
based on accurate the number of customers for
product
based on varieties
accurate (high-volume of to low-volume production) [1]. or solitary, already existing product families analyze the
face-to-face serviceforecasting
industries. Asthe number
part of the of customers
research, for
we are
To cope withservice
face-to-face this augmenting
industries.variety
As partasofwell as to be able
the research, to
we are product structure on a physical level (components level) which
identify possible optimization potentials in the existing causes difficulties regarding an efficient definition and
production
2212-8271 © system,
2017 The it is important
Authors. Publishedtobyhave a precise
Elsevier B.V. knowledge comparison of different product families. Addressing this
Peer-review
2212-8271 ©under
2017responsibility
The Authors. of the scientific
Published committee
by Elsevier B.V.of the 11th CIRP Conference on Intelligent Computation in Manufacturing Engineering.
Peer-review under responsibility of the scientific committee of the 11th CIRP Conference on Intelligent Computation in Manufacturing Engineering.
2212-8271©©2017
2212-8271 2019TheThe Authors.
Authors. Published
Published by Elsevier
by Elsevier B.V. B.V.
Peer-reviewunder
Peer-review underresponsibility
responsibility
of of
thethe scientific
scientific committee
committee of the
of the 28th12th
CIRPCIRP Conference
Design on 2018.
Conference Intelligent Computation in Manufacturing Engineering.
10.1016/j.procir.2019.02.042
680 Takashi Tanizaki et al. / Procedia CIRP 79 (2019) 679–683
T. Tanizaki et al. / Procedia CIRP 00 (2018) 000–000

2.1. Bayesian Linear Regression actual customers during the same period. Table 1 shows
explanatory variables used for forecasting. The forecasting rate
Bayesian Linear Regression (Bayesian) is a method of α, that is ratio of the number of forecastedcustomers to that of
applying Bayesian network to machine learning. The Bayesian actual customers, is calculated using the equations (1) and (2).
network is a probabilistic model in which conditional
dependencies among multiple random variables are expressed Table 1. Explanatory variable
using a graph structure and dependency relationships between Category Explanatory variable Definition
random variables are expressed by conditional probabilities January Jan/1-Jan/31
February Feb/1-Feb/28
[1]. The Bayesian network is defined by three variables:
March Mar/1-Mar/31
random variable, conditional dependency condition between April Apr/1-Apr/30
random variables, and conditional probability [2]. By using the May May/1-May/31
Bayesian network, the probability distribution of unobserved Month
June Jun/1-Jun/30
July Jul/1-Jul/31
variables is calculated using observed some variables and the August Aug/1-Aug/31
value with the highest probability value is obtained as the Septemner Sep/1-Sep/30
predicted value of that variable. October Oct/1-Oct/31
Novenber Nov/1-Nov/30
December Dec/1-Dec/31
2.2. Boosted Decision Tree Regression Monday Weekday and the next day is weekday
Tuesday Weekday and the next day is weekday
Boosted Decision Tree Regression (Boosted) is a method of Wednesday Weekday and the next day is weekday
learning using Boosting [3]. Boosting is machine learning Thuesday Weekday and the next day is weekday
Fryday Weekday and the next day is weekday
using multiple learning devices. In this method, the number of Saturday Even if the target day is a holiday it is Saturday.
learning times of the incorrectly forecasted case is increased in The day of Sunday Sunday and the next day is weekday
order to improve learning accuracy by increasing the weight of the week Even if the target day is a holiday it is Sunday.
Sunday during holidays Sunday and the next day is holiday
that case.
Even if the target day is a holiday it is Sunday.
Holiday Holiday and the nextday is weekday
2.3. Decision Forest Regression Holiday during holidays Holiday and the nextday is holiday
Before holiday Weekday and the next day is holiday
Lastday during holidays The last day of three or more consecutive holidays
Decision Forest Regression (Decision) is a method of
January 1st January 1st
learning using Random Forest [4]. Random Forest is a method January 2nd January 2nd
of constructing a forest using multiple decision trees and January 3rd January 3rd
integrating learning results for each decision tree. Extreme bias Year-end Dec/29-Dec/31
End of year party Weekday of December
in learning of each decision tree can be prevented by Christmas eve December 24
incorporating randomness when extracting learning data to be Event Coming-of-age day Second Monday in January
used in each decision tree. As a result, excessive learning can Setsubun February 2nd
be prevented and high generalization performance can be Obon Aug/13-Aug/15
New year's party Weekday till the coming-of-age day
acquired. except Jan/1-Jan/3
Farewell party Weekday in March
2.4. Stepwise Method Welcome party Weekday in April
Average wind speed Average wind speed per day (m/s)
Maximum wind speed Maximum wind speed per day (m/s)
Stepwise method (Stepwise) is a method of constructing a Highest temperature Highest temperature in a day （℃)
regression model by searching for a combination of objective Lowest temperature Lowest temperature in a day （℃)
variables that can most explain the explanatory variable by Weather Amount of precipitation Amount of precipitation in a day (mm)
Maximum precipitation Maximum amount of precipitation
sequentially increasing or decreasing the objective variable [5].
in ten minutes (mm)
When adding highly objective variables to regression formulas, Maximum instantaneous Maximum instantaneous
there are variables that have already been added, which become wind speed wind speed in a day (m/s)
useless due to their relevance to objective variables added later.
pi : Actual number of customers on i-th day
Therefore, in the stepwise method, each time an objective ei : Forecasting number of customers on i-th day
variable is added, the variable which becomes insignificant for N : Forecast period
the explanatory variable is deleted from the regression formula. αi : Forecasting raito on i-th day
3. Forecasting of customers visiting 𝑝𝑝𝑖𝑖 − |𝑝𝑝𝑖𝑖 − 𝛼𝛼𝑖𝑖 |
𝛼𝛼𝑖𝑖 = (1)
𝑝𝑝𝑖𝑖
3.1. Target data
∑𝑁𝑁
𝑖𝑖=1 𝛼𝛼𝑖𝑖
We got visitor data for 5 stores from the restaurant chain R α= (2)
of the joint researches and forecasted the number of customers 𝑁𝑁
by the method described in Chapter 2. Based on the visit record
of '14/5/1 to' 15/4/30, we forecasted the number of customers
from '15 / 5/1 to '16/4/30 and compared it with the number of
 Takashi Tanizaki et al. / Procedia CIRP 79 (2019) 679–683 681
T. Tanizaki et al. / Procedia CIRP 00 (2018) 000–000

3.2. Forecasting result the number of reserved customers in stores excluding store A
is about 35% to 40%.
Table 2 shows the forecasting results. In the machine
learning, the data usage rate, that is ratio of utilized data for Table 2. Forecasting results
machine learning to total data, is changed from 40% to 100%. Store Data usage Bayesian Boosted Decision Stepwise
Since the Stepwise method is a method using all data, the data rate
usage rate is 100%. The columns of yellow background color A 40% 82.0% 77.4% 80.8% 81.5%
are the result with the highest forecasting rate among each 50% 82.3% 76.7% 80.1%
method. The black box surrounding column shows the highest 60% 82.6% 75.8% 79.7%
forecasting rate among the four methods. For store A, the 70% 82.7% 76.1% 81.4%
forecasting rate using Bayesian is the highest. For stores B and 80% 82.5% 78.2% 80.9%
C, the forecasting rate using Stepwise is the highest. For stores 90% 82.7% 76.1% 81.4%
D and E, the forecasting rate using Decision method is the 100% 82.6% 77.2% 80.7%
highest. The forecasting rate using Boosted tends to be low. In B 40% 80.7% 78.1% 79.2% 81.9%
Bayesian, the higher the data usage rate, the higher the 50% 78.8% 69.2% 76.4%
forecasting rate tends to be. In Boosted and Decision, there is 60% 79.6% 67.2% 78.3%
no noticeable relationship between data usage rate and 70% 80.0% 69.8% 79.5%
forecasting rate. The stores A and B are located in Japan's 80% 80.3% 70.2% 77.8%
leading high-class shopping street adjacent to the business area 90% 80.4% 72.4% 79.8%
and the store E is located adjacent to the central station in the 100% 81.0% 73.8% 80.2%
suburban residential area. C 40% 79.0% 70.9% 76.3% 80.2%
Therefore, many of customers at stores A and B are business 50% 79.1% 68.6% 77.7%
customers and tourists, and many of customers at store E are 60% 79.3% 76.9% 77.1%
residents in the surrounding area. On the other hand, stores C 70% 79.4% 76.6% 79.0%
and D are located in the downtown area adjacent to the central 80% 79.4% 77.0% 78.1%
station with many passengers. Therefore, many of customers at 90% 79.4% 76.2% 77.8%
stores C and D stores are unspecified customers. 100% 79.5% 76.7% 78.8%
Fig. 1 shows the daily forecasting results using Bayesian at D 40% 75.8% 73.0% 78.4% 78.1%
store A with the highest forecasting rate (90%). The forecasting 50% 76.6% 75.0% 75.2%
value and the actual value are in the same trend. The difference 60% 77.7% 72.5% 77.8%
between the average forecasting value and the average actual 70% 77.8% 73.5% 77.4%
value is 15.5, and the difference between the annual total of the 80% 77.9% 74.9% 75.9%
forecasting value and that of the actual value is about 5% of the 90% 77.8% 74.6% 77.6%
actual value. The difference between the forecasting value and 100% 78.3% 75.4% 76.7%
the actual value is large at the Obon , the New Year's Holiday, E 40% 82.8% 81.9% 82.1% 83.3%
the beginning of the fiscal year. 50% 82.6% 80.8% 83.9%
Fig. 2 shows a scatter diagram of "the number of customers 60% 83.3% 81.1% 83.2%
and the number of reserved customers". Since R2 value is 0.75, 70% 83.4% 80.4% 81.2%
there is correlation between these two variables. Table 3 shows 80% 83.3% 79.1% 82.0%
R2-value of the number of customers and the number of 90% 83.5% 79.7% 82.2%
reserved customers and the ratio of the number of reserved 100% 83.9% 80.1% 82.0%
customers to the number of customers. Since the R2 value is 0.6
to 0.7, there is a slight correlation between the number of
customers and the number of reserved customers. The ratio of

Fig. 2. Scatter diagram of the number of customers and reserved customers

Fig. 1. Daily forecasting results

682 Takashi Tanizaki et al. / Procedia CIRP 79 (2019) 679–683
T. Tanizaki et al. / Procedia CIRP 00 (2018) 000–000

Table 3. R2-value and ratio of reserved customers Table 4. Forecasting results of improvement method
Store R2-value Ratio Store Data usage Bayesian Boosted Decision Stepwise
A 0.7518 55.5% rate
B 0.5931 39.3% A 40% 91.2% 89.9% 90.9% 91.8%
C 0.6194 34.1% 50% 91.2% 89.5% 90.4%
D 0.6809 34.6% 60% 91.0% 89.6% 89.8%
E 0.6389 37.2% 70% 91.2% 89.3% 90.6%
80% 91.4% 89.3% 91.2%
3.3. Improvement of forecasting method 90% 91.5% 90.2% 91.4%
100% 91.7% 89.2% 91.0%
From 3.2, it is found that there is slight correlation between B 40% 87.2% 86.2% 87.2% 88.9%
"the number of customer" and "the number of reserved 50% 87.0% 86.3% 87.2%
customers". The ratio of the number of reserved customers is 60% 87.1% 86.1% 86.8%
35% to 40%, which is not low. The motivation for making 70% 87.3% 86.3% 86.9%
reservations at the restaurant is often based on business reasons 80% 87.3% 86.7% 86.9%
such as "business trip" and "dinner meeting" and personal 90% 87.4% 86.7% 86.8%
reasons such as "family anniversary". Although it is possible to 100% 87.6% 87.0% 86.5%
forecast the approximate time when the above events occur, it C 40% 84.6% 83.1% 83.3% 86.0%
is difficult to forecast exactly on a daily basis. On the other 50% 84.5% 83.7% 85.0%
hand, it is necessary to forecast the number of customers 60% 84.7% 84.6% 83.9%
visiting with high accuracy on a daily basis to automate 70% 84.5% 83.8% 84.4%
employees’ work arrangement and food materials ordering, 80% 84.7% 83.8% 83.6%
90% 84.4% 84.2% 84.8%
which is the research goal, Therefore we decide to use the
100% 84.4% 82.9% 84.4%
number of reserved customers that can be grasped from several
days before in this restaurant chain for forecasting the number D 40% 83.8% 83.3% 84.8% 85.7%
of customers. We forecast only the number of customers 50% 84.5% 84.7% 84.2%
without reservation and calculate the forecasting value by 60% 85.1% 83.1% 85.5%
equation (3). 70% 85.0% 83.0% 84.9%
80% 85.1% 82.7% 84.9%
Zi : Total number of forecasting customers on i-th day
90% 85.5% 83.5% 85.1%
Xi : The number of forecasting customers without reservation
on i-th day 100% 85.8% 82.7% 84.2%
Yi : The number of reserved customers on i-th day E 40% 85.2% 86.3% 86.3% 84.6%
50% 84.1% 86.2% 86.0%
Zi = Xi + Yi (3) 60% 84.5% 86.1% 85.0%
70% 84.8% 86.2% 84.0%
Table 3 shows the forecasting results. The column of yellow 80% 84.7% 87.2% 84.5%
background color and is the result with the highest forecasting 90% 84.8% 86.8% 84.2%
rate among each method. The black box surrounding column 100% 85.0% 87.3% 85.5%
shows the highest forecasting rate among the four methods.
The columns with a background color of pink are the highest
forecasting rate in Table 2. The forecasting rate exceeded
approximately 85% in all stores.
Fig. 3 shows the daily forecasting results using Bayesian at
store A with the highest forecasting rate (100%). The graphs of
the forecasting value and the actual value are almost
overlapping. The difference between the average forecasting
value and the average actual value is 0.2, and the difference
between the annual total of the forecasting value and that of the
actual value is about 0.1% of the actual value.
We got the evaluation that this method is practically
applicable from the restaurant R of the joint researches. In the
future, we plan to improve forecasting accuracy and research Fig.3. Daily forecasting results
on the efficiency of store management such as automated food
materials ordering and employees' work arrangement based on
forecasting results. 4. Conclusions

In this paper, demand forecasting methods using internal

data such as POS data and external data in the ubiquitous
environment such as weather, events, etc. are proposed. We use
 Takashi Tanizaki et al. / Procedia CIRP 79 (2019) 679–683 683
T. Tanizaki et al. / Procedia CIRP 00 (2018) 000–000

Bayesian Linear Regression, Boosted Decision Tree Acknowledgements

Regression, Decision Forest Regression and Stepwise method
as the demand forecasting method. There was no big difference
in the forecasting rate using the method of Bayesian, Decision,
and Stepwise, and the forecasting rate of Boosted was a little
low. The forecast rate of any store exceeded approximately
85%.
We got the evaluation that this method is practically
applicable from the restaurant R. In the future, we plan to
improve forecasting accuracy and research on the efficiency of
store management such as automated food materials ordering
and employees' work arrangement based on forecasting results.
This study is supported by JSPS KAKENHI (16H02909).
References
[1] Motomura Y. Baysian network. Technical Report of IEICE 2003; 103-285:
25-30 (In Japanese).
[2] Motomura Y. Baysian Network Softwares. Journal of the Japanese Society
for Artificial Intelligence 2002; 17-5: 1-6 (In Japanese)
[3] Freund Y, Schapire R, Abe N (translation). A short Introduction to Boosting,
Journal of the Japanese Society for Artificial Intelligence 1999; 14-5: 771-
779 (In Japanese).
[4] Habe H. Random Forests, ISPJ SIG Technical Report 2002; 2012-CVIM-
182-31: 1-8 (In Japanese).
[5] Bolch BW, Huang CJ, Nakamura K (translation). Applied Statistics
analysis, Tokyo: Morikita Publishing Co., Ltd.; 1968 (In Japanese).