Research ArticleComparison of Two Quantitative Analysis Techniques toPredict the Evaluation of Product Form Design
Hung-Yuan Chen,1 Yu-Ming Chang,2 and Ting-Chun Tung1
1 Department of Visual Communication Design, Southern Taiwan University of Science and Technology, Tainan 71005, Taiwan2Department of Creative Product Design, Southern Taiwan University of Science and Technology, Tainan 71005, Taiwan
Correspondence should be addressed to Hung-Yuan Chen; [email protected]
Received 12 June 2014; Accepted 4 September 2014; Published 21 September 2014
Academic Editor: Teen-Hang Meen
Copyright Β© 2014 Hung-Yuan Chen et al. This is an open access article distributed under the Creative Commons AttributionLicense, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properlycited.
Consumer satisfaction with a productβs form plays an essential role in determining the likelihood of its commercial success. Aconsumer perception-centered design approach is proposed in this study to aid product designers with incorporating consumersβperceptions of product forms in the design process. The consumer perception-centered design approach uses the linear modelingtechnique (multiple linear regression) and the nonlinear modeling technique (neural network) to determine the satisfying productform design formatching a given product image. A series of experimental evaluations are conducted to collect evaluation results forexamining the relationship between the automobile profile features and the consumersβ perceptions of the automobile image. Theresult of predictive performance comparison shows that both the nonlinear neural network modeling technique and the multiplelinear regression technique are comparably good for predicting the consumersβ likely response to a particular automobile profilesince the predictive performance difference between the two modeling techniques is very slight in this study. Although this studyhas chosen a 2D automobile profile for illustration purposes, the concept of the proposed approach is expansively applicable to 3Dautomotive form design or other consumer product forms.
1. Introduction
Consumers interact with a huge number of diverse productsduring the course of their daily lives and therefore sub-consciously develop sophisticated product evaluation skills.A consumerβs purchase decision is based not only on aproductβs functionality and fitness for use, but also on thepsychological response induced by its physical appearance.This phenomenon is particularly apparent in the case ofmature consumer products such as automobiles, mobilephone, tableware, and computer accessories. These matureconsumer products can still sell well in the market placeeven when lacking advanced technological features andfunctionalities provided that their form design finds favorwith the consumers. As a result, consumer satisfaction witha productβs form plays an essential role in determiningthe likelihood of its commercial success. However, productform design activities are often reduced to a discussion
based on the designersβ opinions and personal subjectivities,with no theoretical basis. To avoid the subjective judgmentsin the design process and to objectively relate consumersβpsychological satisfaction with a product to its form features,many systematic design approaches have been proposedfor modeling the correlation between the form features ofa product and the consumersβ perception of the productimage [1β4]. Amongst such techniques, Kansei Engineering(KE) is a fundamental consumer-oriented systematic designapproach in which the consumersβ feelings or product imageperceptions are expressed using suitable image descriptors.KE has emerged as one of the most powerful techniquesfor taking account of the correlation between a productβsattributes and the induced product image during the designprocess [5].
The effectiveness of the KE approaches is crucially deter-mined by the choice of analytical technique with which onecan model the correlation between the product form and
Hindawi Publishing CorporationMathematical Problems in EngineeringVolume 2014, Article ID 989382, 9 pageshttp://dx.doi.org/10.1155/2014/989382
2 Mathematical Problems in Engineering
the corresponding consumer perception. The models usedin the product design field to predict the likely consumerresponse to a particular product form are commonly basedon either the conventional linear analysis techniques orthe nonlinear modeling techniques (such as the artificialintelligent system). Conventional linear analysis techniquessuch as multiple linear regression (MLR) [6] and quantitativetheory type I [4, 7] are commonly employed to interpretthe relationships between the independent and dependentvariables. MLR and quantitative theory type I are widelyused because they are easy, are simple, and have goodpredictive performance. MLR is particularly good when theinput data and output data relationship is linear. However,MLR and quantitative theory type I do not properly handlenonlinear relationships very well since the accuracy of thepredicted results is seriously degraded if the independentand dependent variables are characterized by a nonlinearrelationship. In contrast to conventional linear analysis tech-niques, nonlinear modeling techniques are defined as anemerging approach to learning the humanmind in an uncer-tainty environment [8] and are free of the restriction on thetype of relationship between the independent variables anddependent variables when constructing prediction models.Therefore, are the nonlinear modeling techniques suitable forexploring the relationship between the product form featuresand the corresponding consumersβ perceptions? Or are thelinear modeling techniques good enough to do so [9]? Whatkind of modeling technique should be used to predict thelikely consumer response to a particular product form? Toillustrate how themodeling techniques can be used to answerthese research questions, this study considers the design ofan automobile profile and explores the correlations betweenthe design variables of automobile profile and the associatedimage perceptions using a linear analysis technique and anonlinear modeling technique, respectively. This particulardesign case is chosen because the lateral contours of avehicle are known to supply powerful stimuli influencingthe consumerβs image perceptions and the distinctivenessof a vehicleβs outline is receiving increasing emphasis onmanufacturersβ marketing strategies nowadays.
Multiple linear regression (MLR) is a technique formodeling and analyzing numerical data consisting of thevalues of a large number of independent variables and asmaller number of dependent variables. MLR enables theidentification of a set of independent variables which explaina certain proportion of the variance in a dependent variableat a specified significance level. The relationship betweenthe independent variables and the dependent variables canbe illustrated either graphically or, more usually, by meansof an equation (model). Having established this model, itcan be used to predict the value of the dependent variablefor any given set of independent variables. Because of theeffective learning and prediction capabilities for analyzingthe relationship between the product form features (theinput variables) and the consumersβ perceptions (the out-put variables), nonlinear neural networks (NNs) have beensuccessfully applied in a diverse range of fields [10, 11]. NNsare well suited to formulate the product design process formatching the product form to the consumersβ perceptions,
which is often a black box and cannot be precisely described[10, 12, 13].
From the discussions above, a consumer perception-centered design approach is proposed for modeling thecorrelation between the profile features of an automobile andthe consumersβ perception of the image projected by the auto-mobile. Consumersβ perceptions of the automobile profileimage are described using single adjectives and a coordinate-based definition is used to define the automobile profile.In order to gather the evaluation data of consumers, theconsumersβ perception evaluations of automobile image areconducted.The predictionmodels of consumersβ perceptionsinduced by the automobile profile images are constructedusing MLR and NNs. The subsequent sections of this studyare organized as follows. Section 2 presents a review of MLRand NNs. Section 3 presents the research implementation.Section 4 analyzes the results of constructing predictionmodels. Section 5 verifies the performance of predictionmodels and provides the predictive performance comparisonbetween MLRmodels and NNmodels. Section 6 offers somebrief conclusions.
2. Review of Quantitative Analysis Techniques
In this section, we present a brief outline of the relevant the-ories and algorithms, including the MLR and the NNs.Thesetechniques are used to examine the relationship between theautomobile profiles and the corresponding product images inthis study.
2.1. Multiple Linear Regression (MLR). MLR is a traditionalstatistical technique for modeling the linear relationshipsbetween the input variables (i.e., the independent variables)π₯πβ π π and the desired output variables (i.e., the dependent
variables) π¦π
β π . The regression equation for an MLRproblem involving π data samples has the form
π¦π= π0+ π1π₯π,1
+ β β β + πππ₯π,π
+ ππ, π = 1, . . . , π, (1)
where π0is the regression constant, π
1, . . . , π
πare the partial
regression coefficients corresponding to the π input variables,and π
πis the error term. The regression coefficients are
generally estimated using an ordinary least squares (OLS)procedure such that the sum of the squared errors (βπ
π=1ππ)
is minimized. In MLR, the coefficient of determination, π 2,indicates the percentage of the variation in π¦
πexplained
by the independent variables π₯π. In other words, the value
of π 2 (ranging from 0 to 1) indicates the goodness of
fit of the regression model. The coefficients π1, . . . , π
πin
(10), commonly referred to as the unstandardized regressioncoefficients, can be used to construct the regression equationdirectly. However, the standardized regression coefficients,π½1, . . . , π½
π, provide a more suitable means of analyzing the
relative importance of the different variables. Note that eachstandardized regression coefficient represents the change inresponse of appropriate per standard deviation change in oneoutput value π¦.
Having established this equation (model), it can be usedto predict the value of the dependent variable for any given
Mathematical Problems in Engineering 3
set of independent variables. In addition, as described inthe following discussions, MLR enables a screening of theindependent variables such that the performance of theprediction model is enhanced. It is known that increasingthe number of independent variables in a regression equationimproves the fit to the training set but reduces the predictiveability of the model when supplied with a set of input datataken from outside the training set. Therefore, it is neces-sary to identify the subset of independent variables whichcollectively enhance the accuracy of the prediction model.Generally speaking, the optimal set of independent variablesis compiled using one of three different methods, namely, theforward selectionmethod, the backward eliminationmethod,or the stepwise method [14]. Of these three methods, MLRwith a stepwise procedure is particularly advantageous sinceit provides the combinatorial benefits of the forward selectionmethod and the backward elimination method, respectively.
2.2. Neural Networks. NNs are nonlinear models and arewidely used to examine the complex relationship betweeninput variables and output variables. In this study, we use themultilayered feed-forward neural networks trained with theback-propagation learning algorithm, as it is an effective andthe popular supervised learning algorithm. A typical three-layer network consists of an input layer, an output layer, andone hidden layer, with π, π, and π neurons, respectively(indexed by π, π, and π, resp.) [15]. π€
ππand π€
ππrepresent the
weights for the connection between neuron π (π = 1, 2, . . . , π)
and neuron π (π = 1, 2, . . . , π) and between neuron π (π =
1, 2, . . . , π) and neuron π (π = 1, 2, . . . , π), respectively.In training the network, a set of input patterns or signals,(π₯1, π₯2, . . . , π₯
π), is presented to the network input layer. The
network then propagates the inputs from layer to layer untilthe output layer generates the outputs. This involves thegeneration of the outputs (π¦
π) of the neurons in the hidden
layer as given in (2) and the outputs (π¦π) of the neurons in
the output layer as given in (3). One has
π¦π= π(
π
β
π=1
π₯ππ€ππβ ππ) , (2)
π¦π= π(
π
β
π=1
π₯ππ€ππ
β ππ) , (3)
where π(β ) is the sigmoid activation function as given in (4)and π and π are threshold values
π (π) =
1
1 + πβπ
. (4)
If the outputs (π¦π) generated by (3) are different from the
target outputs (π¦βπ), errors (π
1, π2, . . . , π
π) are calculated by (5)
and then propagated backwards from the output layer to theinput layer in order to update the weights for reducing theerrors:
ππ= π¦β
πβ π¦π. (5)
The weights (π€ππ) at the output neurons are updated as π€
ππ+
Ξπ€ππ, where Ξπ€
ππis computed by (known as the delta rule)
Ξπ€ππ
= πΌπ¦ππΏπ, (6)
where πΌ is the learning rate (usually 0 < πΌ β€ 1) and πΏπis the
error gradient at neuron π, given as
πΏπ= π¦π(1 β π¦
π) ππ. (7)
The weights (π€ππ) at the hidden neurons are updated as π€
ππ+
Ξπ€ππ, where Ξπ€
ππis calculated by
Ξπ€ππ= πΌπ₯ππΏπ, (8)
where πΌ is the learning rate (usually 0 < πΌ β€ 1) and πΏπis the
error gradient at neuron π, given as
πΏπ= π¦π(1 β π¦
π)
π
β
π=1
πΏππ€ππ. (9)
The training process is repeated until a specified errorcriterion is satisfied.
3. Implementation Procedure and Steps
3.1. Automobile Profile Definition and Evaluation Samples.This study collected a large number of pictures of commercialautomobiles and the corresponding body length datum.These pictures were then examined to identify the variousprofile characteristics which collectively define the overallautomobile profile based on a human interpretation of thedistinctive component features of that particular profile. Atotal of 144 vehicle pictures with a side-view orientation werecollected from automobile magazines, catalogs, and websites.The 144 automobile images were then converted into profilesamples by tracing their contour features with 16 Beziersegments using computer graphic software.
Table 1 shows a typical example of a car profile con-structed using the Bezier segments. Of the 16 Bezier segments(πΆ1βΌ πΆ16) in the profile, the chassis segment (πΆ
15) was
assumed to be a straight line with just two control pointssince this profile feature is approximately straight in mostautomobiles. However, the remaining segments were eachassigned four control points to enable the introduction ofdetailed profile variations in each segment. As stated above,each profile was constructed using 16 Bezier segments. As aresult, a total of 16 control points were located at the jointsbetween contiguous Bezier curves. Since these joints wereused by both segments when tracing the automobile profile,the overall car profile was defined using a total of 46 controlpoints (i.e., 2 + (15 Γ 4) β 16 = 46). In order to define theautomobile profile, a coordinate-based definition approachwas conducted in this study. Control point P
1was specified
as the origin (0, 0) and the coordinate positions of each of theother control points were then recorded with respect to thisorigin point. Since each control point was assigned both π-and π-axis coordinates, the automobile profile was definedusing a total of 92 (i.e., 46Γ2) coordinate data items. Further,
4 Mathematical Problems in Engineering
Table 1: Tracing of automobile profile using 16 Bezier curves.
Origin(0, 0)
P44P43
P40
P37
P34
P31
P28P25
P22P19
P16
P13
P10
P4
P6
P1
C1
C2
C3
C4
C5 C6 C7 C8 C9
C10
C11
C12
C13
C14
C15
C16
Curve segment designation Point Design variablesπΆ1: under-part segment of fore bumper π
1βΌ π4
π1(π)
βΌ π8(π)
πΆ2: middle-part segment of fore bumper π
4βΌ π7
π7(π)
βΌ π14(π)
πΆ3: upper-part segment of fore bumper π
7βΌ π10
π13(π)
βΌ π20(π)
πΆ4: fore lamp segment π
10βΌ π13
π19(π)
βΌ π26(π)
πΆ5: engine hood segment π
13βΌ π16
π25(π)
βΌ π32(π)
πΆ6: fore windshield segment π
16βΌ π19
π31(π)
βΌ π38(π)
πΆ7: car roof segment π
19βΌ π22
π37(π)
βΌ π44(π)
πΆ8: rear windshield segment π
22βΌ π25
π43(π)
βΌ π50(π)
πΆ9: trunk segment π
25βΌ π28
π49(π)
βΌ π56(π)
πΆ10: rear lamp segment π
28βΌ π31
π55(π)
βΌ π62(π)
πΆ11: upper-part segment of rear bumper π
31βΌ π34
π61(π)
βΌ π68(π)
πΆ12: middle-part segment of rear bumper π
34βΌ π37
π67(π)
βΌ π74(π)
πΆ13: upper-part segment of rear bumper π
37βΌ π40
π73(π)
βΌ π80(π)
πΆ14: protection hood segment of rear wheel π
40βΌ π43
π79(π)
βΌ π86(π)
πΆ15: chassis segment π
43βΌ π44
π85(π)
βΌ π88(π)
πΆ16: protection hood segment of front wheel π
44βΌ π1
π87(π)
βΌ π2(π)
each automobile profile samplewas painted 50%gray in orderto enhance its visual impact and was uniformly reduced insize using a scaling factor of 1 : 25 based on actual automobilebody length to individually display on an A4-sized card. Thetotal set of 144 automobile profile samples were considered inthe subsequent evaluation trials.
3.2. Selecting the Representative Image Descriptors of Auto-mobile Profile. Consumers commonly use image words (i.e.,adjectives) to express their image perceptions of a product.Althoughmany different imagewords are usedwhen describ-ing everyday products, the image words applicable to anautomobile profile are more limited. In this study, 3 designersand 3 white-collar individuals from nondesign backgroundswere invited to participate in a discussion aimed at identifyingsuitable descriptors with which one can describe the possiblepsychological responses of a consumer when presented withthe sample automobile profiles. The image descriptors wereelicited from the participants using the following four-stepprocedure.
Step 1. The 144 automobile profiles were reviewed, and theimage words used by the individual participants to describetheir perceptions of the image projected by each automobileprofile were recorded.
Table 2: Selection and classification of image descriptors.
Imagedescriptors Image words contained
Modern β Modern, Advanced, Technical, Novel, Fashionable
Formal β
Formal, Popular, Robust, Geometric, Rational,Mature
Classical β Classical, Sleek, Stylish, Smooth
Rakish β
Rakish, Personal, Dashing, Young, Peculiar,Speedy, Future, Streamlined, Wild
Elegant β
Elegant, Gorgeous, Luxurious, Noble, Lofty,Gentle
Family β Family, Leisure, Vital
Step 2. The focus group method [16] was applied to select 33suitable product image descriptors for a generic automobileprofile.
Step 3. The Kawakita Jiro (K J) method [17] was thenapplied to classify the 33 image words in accordance withtheir semantic similarities. As shown in Table 2, six basicdescriptor groups were identified.
Step 4. From each group, one image descriptor was chosento represent the overall characteristics of the group, that is,Modern, Formal, Classical, Rakish, Elegant, and Family.
Mathematical Problems in Engineering 5
Table 3: Results of MLR analysis in six image perception domains.
MLR prediction models π π 2
βRakishβ image= β0.123(V.3π₯) β 0.138(V.9π₯) + 0194(V.17π₯) β 0.549(V.19π₯) + 0.400(V.21π₯) + 0.182(V.31π₯)+ 0.171(V.39π₯) + 0.112(V.42π¦) β 0.107(V.43π₯) β 0.523(V.44π¦) + 0.105(V.49π₯) + 0.377(V.50π¦)β 0.087(V.58π¦) + 0.162(V.60π¦) β 0.184(V.67π₯) β 0.060(V.73π₯) β 0.335(V.88π¦) + 0.263(V.89π₯)+ 4.450
0.917 0.819
βFamilyβ image= 0.067(V.3π₯) β 0.098(V.5π₯) + 0.092(V.9π₯) + 0.099(V.11π₯) + 0.338(V.19π₯) β 0.382(V.21π₯)β 0.313(V.30π¦) β 0.286(V.38π¦) + 0.098(V.42π¦) + 0.541(V.44π¦) + 0.145(V.48π¦) + 0.053(V.51π₯)+ 0.204(V.58π¦) β 0.265(V.76π¦) β 0.203(V.82π¦) + 5.804
0.840 0.671
βFormalβ image= 0.075(V.7π₯) β 0.037(V.15π₯) β 0.078(V.17π₯) β 0.390(V.32π¦) + 0.098(V.33π₯) + 0.386(V.36π¦)β 0.113(V.37π₯) + 0.095(V.56π¦) + 0.041(V.57π₯) β 0.124(V.63π₯) + 0.112(V.68π¦) + 0.209(V.69π₯)β 0.279(V.70π¦) + 0.146(V.78π¦) β 0.165(V.87π₯) β 0.266(V.92π¦) + 5.435
0.849 0.685
βClassicalβ image= β0.043(V.7π₯) + 0.128(V.9π₯) β 0.127(V.10π¦) β 0.038(V.15π₯) β 0.094(V.23π₯) + 0.101(V.33π₯)β 0.212(V.34π¦) + 0.195(V.36π¦) β 0.140(V.37π₯) + 0.084(V.41π₯) + 0.171(V.48π¦) β 0.140(V.56π¦)+ 0.022(V.59π₯) + 0.040(V.73π₯) β 0.039(V.77π₯) β 0.085(V.92π¦) + 5.208
0.904 0.791
βElegantβ image= β0.067(V.3π₯) β 0.228(V.18π¦) + 0.522(V.20π¦) β 0.308(V.22π¦) β 0.075(V.25π₯) + 0.364(V.44π¦)β 0.046(V.47π₯) + 0.121(V.71π₯) β 0.214(V.76π¦) + 5.259
0.813 0.639
βModernβ image= 0.050(V.5π₯) β 0.128(V.9π₯) β 0.201(V.10π¦) + 0.165(V.12π¦) + 0.100(V.25π₯) + 0.160(V.39π₯)β 0.072(V.41π₯) + 0.072(V.42π¦) β 0.120(V.43π₯) β 0.517(V.44π¦) + 0.202(V.48π¦) + 0.103(V.49π₯)β 0.157(V.57π₯) + 0.193(V.61π₯) + 0.174(V.62π¦) β 0.060(V.73π₯) β 0.151(V.84π¦) + 0.217(V.92π¦)+ 4.304
0.836 0.661
3.3. Evaluation of Automobile Profile Images. The 144 auto-mobile profile samples and the six image descriptors wereused to perform an examination of the correlation betweenthe automobile profiles and their associated image percep-tions. The investigation was performed by 32 subjects. Inthe evaluation process, each profile image was assessed interms of the six image descriptors. Note that the evaluationmethod was performed separately by each of the 32 subjects.Taking the first image descriptor, each subject divided the144 automobile profile samples into 3 groups, that is, low(L), medium (M), and high (H), in accordance with his orher intuitive perception of the extent to which the imagedescriptor described the feelings induced by the profile. Thesubject then further divided each of the 3 groups into 3subgroups. In this way, the 144 automobile profile sampleswere divided into a total of 9 groups. Each automobile profilewas then assigned a score from 1 to 9 (LL to HH) dependingon the group to which it was assigned. Having classified the144 profiles in terms of the first image descriptor, the two-stage classification process was repeated for each of the otherfive image descriptors. Finally, the evaluation results obtainedfrom the 32 subjects for the 144 automobile profile sampleswere used to analyze the relationship between the automobileprofile features and the associated image perceptions.
4. Constructing Prediction Models ofAutomobile Profile Image
4.1. MLR Prediction Models. MLR with a stepwise selectionprocedure was applied to the evaluation data obtained for
the automobile profiles in order to construct functionalrelationships between the design variables of the automobileprofile and the corresponding consumer responses in eachof the image perception domains. The MLR models wereconstructed subject to the criteria that a π value of lessthan 0.1 was required for entry to the model such that onlythose design variables having a significant effect on the con-sumersβ perception of the automobile profile were retained.In performing the MLR analyses, the independent variablescorresponded to the 92 design variables of automobile profileand the six dependent variables were specified as the meanvalue of the evaluation scores assigned in the correspondingimage perception domain.
Overall, Table 3 shows the functional models relating thedesign variables of the automobile profile to an evaluativerating in each of the six product image perception domains.In this table, themultiple correlation coefficientπ varies from0.813 (Elegant) to 0.917 (Rakish) while the adjusted π
2 valuesvary from 0.639 to 0.819, respectively. This result implies theexistence of a significant relationship between the 92 designvariables and the six image perception domains.
4.2. NN Prediction Models. In constructing and trainingthe NN used to predict the relationship between the 144automobile profiles and the corresponding image descriptors,the 92 design variables for each automobile profile wereused as the input neurons of the NN and the averagevalues of the 6 image evaluations were used as the targetvalues of the output neurons. In attempting to establish theoptimal NN model, three different models were constructed,
6 Mathematical Problems in Engineering
Table 4: RMSE of NN prediction models for automobile profiletraining dataset.
Number oftrainingepochs
NN-Arith. NN-Geom. NN-SumI: 92 neuronsH: 49 neuronsO: 6 neurons
I: 92 neuronsH: 23 neuronsO: 6 neurons
I: 92 neuronsH: 98 neuronsO: 6 neurons
1000 0.1028 0.1031 0.10692000 0.0956 0.0970 0.09645000 0.0846 0.0871 0.087210,000 0.0775 0.0787 0.079120,000 0.0672 0.0685 0.068430,000 0.0620 0.0638 0.064340,000 0.0588 0.0606 0.061050,000 0.0576 0.0599 0.060360,000 0.0567 0.0587 0.059270,000 0.0558 0.0577 0.058480,000 0.0553 0.0570 0.057990,000 0.0550 0.0570 0.0577βI,β βH,β and βOβ indicate the input, hidden, and output layer, respectively.
differing only in terms of the number of neurons withintheir single hidden layer. The number of neurons in thehidden layer was specified in accordance with three commonrules, namely, the arithmetic mean (Arith.), the geometricmean (Geom.), and the sum (Sum) of the input and outputneurons, respectively. As stated above, the 144 automobileprofiles and their associated product images were defined interms of the 92 coordinate variables and 6 product imagedescriptors, respectively. Consequently, each of the three NNmodels comprised 92 input neurons and 6 output neurons. Inaccordance with the three rules specified above, the hiddenlayers of the three models therefore comprised 49 (Arith), 23(Geom.), and 98 (Sum) neurons, respectively. The three NNmodels were trained using the coordinate-based definitiondata associated with the 144 automobile profile samples andthe corresponding image perception values obtained from theevaluation trials. The three NN models were trained usinga sigmoid transformation function and a delta-rule learningrule, respectively.
As shown inTable 4, all threemodels converged rapidly toan RMSE value of just over 0.1 after 1000 training epochs andto a value of just less than 0.1 after approximately 2000 epochs.These results indicate that the NN models all successfullyconverged as the number of training epochs increased. Thethree NN models were trained continuously for a total of90,000 epochs. As shown in Table 4, the NN-Arith. modelobtained its best predictive performance after 90,000 epochs(RMSE = 0.0550), while the NN-Geom and NN-Summodels obtained their best performances after 80,000 epochs(RMSE = 0.0570) and 90,000 epochs (RMSE = 0.0577),respectively. Although the difference between the RMSEvalues of the three NN models is less than 0.0030 (0.0577 β
0.0550 = 0.0027), the RMSE value of the NN-Arith. modelis lower (i.e., better) than those of the other two models after90,000 training epochs. Consequently, the NN-Arith. model
was adopted as the operational NN model for predicting thevalues of each of the six automobile image descriptors for anygiven automobile profile definition data.
5. Validation and Comparison of PredictionModel Performance
5.1. Performance Evaluation of Prediction Models. To verifythe predictive ability of the MLR and NN models, sixnew automobile profiles were designed using the respectivecoordinate-based definitions. The six automobile profileswere displayed on individual A4-sized cards and were evalu-ated by a group of 30 subjects using nine-point Likert scales.The product image perception in each of the six domainsassociated with the automobile profile was also predicted bysubstituting the relevant design variable values of each veri-fication sample into the functional models. The discrepancybetween the Likert scale evaluation scores assigned to eachverification sample by the 30 subjects and those predicted bythe functional models was then assessed using the followingroot-mean-square-error index (EI):
EI = ββπ
π=1(π₯πβ π₯π)2
π
,(10)
where π₯πis the Likert scale point assigned by the πth subject,
π₯πis the Likert scale point predicted by the corresponding
model, and π is the number of participants involved in thevalidation experiments.The EI value in (10) gives the averagedifference between the evaluations of the subjects and thoseof the corresponding model, respectively, for a single pointon the nine-point Likert scale. To enable the reliability ofthe prediction models to be more conveniently compared,a normalized EI value was obtained by dividing the resultobtained from (10) by nine to yield an error rate (ER) inthe interval [0, 1]. Clearly, a lower value of ER indicates anenhanced predictive capability.
Tables 5 and 6 present the predictions, EI values, andcorresponding error rate (ER) values of the MLR and NNmodels, respectively, for the six image perception domainsassociated with the six automobile verification examples (seeFigure 1). From inspection, it is determined that the averageER varies from 13.32% for the βFamilyβ image to 20.57%for the βModernβ image when evaluated using the MLRprediction models (see Table 5). By contrast, the averageER is found to vary from 13.38% for the βFamilyβ imageto 19.22% for the βClassicalβ image when evaluated usingthe NN prediction model (see Table 6). Overall, the resultsdemonstrate the basic reliability of the two functional modelsin predicting the image projected by an automobile profile ineach of the six perception domains.
5.2. Prediction Performance Comparison between MLRModeland NN Model. Table 7 summarizes the average ER valuesof the two functional models when applied to predict theconsumer response to the automobile profile in each ofthe corresponding image perception domains. The onlyexception to this tendency occurs in the βFamilyβ and
Mathematical Problems in Engineering 7
Table 5: Performance evaluation of MLR prediction models.
Mean Prediction ER (EI/9) Mean Prediction ER (EI/9)Rakish Family
1 5.23 5.10 12.8% 7.70 6.59 17.6%2 2.77 2.88 12.8% 6.33 5.81 15.0%3 5.23 4.99 17.1% 7.57 7.20 11.1%4 2.33 2.49 14.3% 8.43 8.49 8.0%5 1.77 4.56 31.9% 6.23 6.56 13.6%6 6.80 7.08 13.3% 6.37 5.97 14.6%
Average ER = 17.03% Average ER = 13.32%Formal Classical
1 4.33 5.07 21.0% 3.40 4.91 22.0%2 6.63 5.42 21.6% 5.53 5.72 13.2%3 5.10 5.01 14.2% 4.53 5.77 23.6%4 6.47 6.41 15.4% 5.40 4.74 18.0%5 4.73 3.27 22.4% 3.20 3.95 12.1%6 3.47 4.89 20.5% 4.93 5.29 20.1%
Average ER = 19.18% Average ER = 18.17%Elegant Modern
1 4.50 4.24 18.1% 4.93 4.14 18.7%2 8.10 6.63 19.0% 3.97 4.10 20.2%3 7.70 6.59 17.6% 5.23 4.84 16.5%4 5.17 7.29 27.1% 4.60 3.91 21.0%5 4.93 4.85 14.9% 4.63 6.25 24.9%6 4.57 5.00 14.2% 5.47 6.39 22.1%
Average ER = 18.48% Average ER = 20.57%
Table 6: Performance evaluation of NN prediction model.
Mean Prediction ER (EI/9) Mean Prediction ER (EI/9)Rakish Rakish
1 5.23 5.71 13.8% 7.70 8.18 13.6%2 2.77 3.73 16.5% 6.33 6.52 14.0%3 5.23 4.17 20.8% 7.57 6.89 12.7%4 2.33 4.12 24.4% 8.43 7.84 10.3%5 1.77 3.02 16.2% 6.23 6.84 14.7%6 6.80 5.72 20.4% 6.37 6.88 15.0%
Average ER = 16.30% Average ER = 13.38%Rakish Rakish
1 4.33 5.13 21.3% 3.40 5.07 23.4%2 6.63 5.97 18.4% 5.53 5.12 13.8%3 5.10 5.37 14.5% 4.53 6.33 27.7%4 6.47 5.97 16.3% 5.40 6.11 18.2%5 4.73 4.37 15.9% 3.20 3.72 12.4%6 3.47 5.03 21.7% 4.93 5.18 19.8%
Average ER = 18.02% Average ER = 19.22%Rakish Rakish
1 4.50 4.23 18.1% 4.93 4.53 17.1%2 8.10 6.99 15.6% 3.97 4.44 20.9%3 7.70 8.02 13.1% 5.23 4.58 17.5%4 5.17 7.66 30.8% 4.60 4.03 20.6%5 4.93 5.67 17.0% 4.63 5.03 17.9%6 4.57 4.81 13.6% 5.47 5.12 20.0%
Average ER = 18.03% Average ER = 19.00%
8 Mathematical Problems in Engineering
v.1 v.2 v.3
v.4 v.5 v.6
Figure 1: Automobile profiles used for functional model verification purposes.
the βClassicalβ image perception domain, in which theaverage ER value of the MLR model (Family = 13.32%;Classical = 18.17%) is slightly lower than that of the NNmodel (Family = 13.38%; Classical = 19.22%). Observingthe mean ER data presented in the lower row of the table, themean average ER values of theMLRmodel and theNNmodelfor the automobile profile are 17.79% and 17.33%, respectively.Comparing the mean average ER value of the MLR modelwith that of the NN model, it is observed that the latter isslightly lower (i.e., better) than the former.This finding seemsto imply that the predictive performance of the NN model isslightly superior to that of the MLR model. However, whilethe inclusion of a large number of design variables enhancesthe efficiency of a general predictive model, it also increasesthe model complexity and therefore causes the designerproblems in understanding the true nature of the relationshipbetween the consumersβ image perception and the individualdesign variables. Although the average ER values of the MLRmodel are slightly inferior to those of the NN, the MLRmodel has the advantage that it can use fewer design variablesto predict the consumersβ likely response to the automobileprofile, and thus the functional relationships between theinput design variables and the image perception values arenot only more straightforward than those in the NN modelbut also more intuitively understandable. From inspection,the difference in the mean average ER values of the MLR andNNmodels is found to be less than 0.5% (17.79%β 17.33% =
0.46%). Thus, overall, it can be inferred that the MLR modelalso represents a good solution for predicting the consumerβsimage perceptions of the automobile profile since it achievesa predictive ability very close to that of the NN model.
6. Conclusions
In this study, the relationships between the independentvariables and the dependent variables are established usingMLR and NNmodeling techniques.The various models haveall been verified by comparing the predicted consumersβperception of the product image in each product imagedomain with the corresponding manual evaluation results.Although the verification result presented in Table 7 hasshown that the predictive performance of the NN model isslightly better than that of the MLR model, the differencein the predictive performance of the two models is very
Table 7: Comparison of average ER values of twopredictionmodels.
MLR predictionModel
NN predictionModel
Rakish 17.03% β16.30%Family β13.32% 13.38%Formal 19.18% β18.02%Classical β18.17% 19.22%Elegant 18.48% β18.03%Modern 20.57% β19.00%Mean of all averageER 17.79% β17.33%
Asterisks indicate that average ER value of prediction model is lower (i.e.,better).
slight here. Thus, this finding implies that the nonlinear NNmodeling technique and the MLR technique are comparablygood for predicting the consumersβ likely response to aparticular automobile profile. However, NNmodel with theirsophisticated nonlinear algorithms is often opaque and itis therefore frequently difficult to recognize the specificdesign variables which dominate the consumersβ responseto the product design. By contrast, MLR technique allowsdesigners to construct relationship models comprising onlythose independent variables which exert the most signifi-cant effect on the dependent values. Nevertheless, nonlinearrelationship modeling using MLR technique results in poorpredictive performance. In future study, it would be worth-while considering the use of the integration model whichcombines the variable selection advantage of MLR and thesophisticated data analysis capabilities of NN for establishingthe relationship between the dependent variables and theindependent variables in the product form design field.Although this study has chosen a 2D automobile profile forillustration purposes, the concept of the proposed approach isexpansively applicable to 3D automotive form design or otherconsumer product forms.
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper.
Mathematical Problems in Engineering 9
References
[1] H. H. Lai, Y. C. Lin, C. H. Yeh, and C. H. Wei, βUser-orienteddesign for the optimal combination onproduct design,β Interna-tional Journal of Production Economics, vol. 100, no. 2, pp. 253β267, 2006.
[2] C. C. Chen andM.C. Chuang, βIntegrating theKanomodel intoa robust design approach to enhance customer satisfaction withproduct design,β International Journal of Production Economics,vol. 114, no. 2, pp. 667β681, 2008.
[3] K.-C.Wang, βA hybrid Kansei engineering design expert systembased on grey system theory and support vector regression,βExpert Systems with Applications, vol. 38, no. 7, pp. 8738β8750,2011.
[4] Y.-C. Lin, C.-H. Yeh, C.-C. Wang, and C.-C. Wei, βIs the linearmodeling technique good enough for optimal form design?A comparison of quantitative analysis models,β The ScientificWorld Journal, vol. 2012, Article ID 689842, 13 pages, 2012.
[5] M. Nagamachi, βKansei engineering as a powerful consumer-oriented technology for product development,β AppliedErgonomics, vol. 33, no. 3, pp. 289β294, 2002.
[6] Y.-M. Chang and H.-Y. Chen, βApplication of novel numericaldefinition-based systematic approach (NDSA) to the designof knife forms,β Journal of the Chinese Institute of IndustrialEngineers, vol. 25, no. 2, pp. 148β161, 2008.
[7] Y. C. Lin, H. H. Lai, and C. H. Yeh, βConsumer-orientedproduct formdesign based on fuzzy logic:A case study ofmobilephones,β International Journal of Industrial Ergonomics, vol. 37,no. 6, pp. 531β543, 2007.
[8] P. T. Helo, Q. L. Xu, S. J. Kyllonen, and R. J. Jiao, βIntegratedVehicle Configuration System-Connecting the domains ofmasscustomization,β Computers in Industry, vol. 61, no. 1, pp. 44β52,2010.
[9] R. B. Page andA. J. Stromberg, βLinearmethods for analysis andquality control of relative expression ratios from quantitativereal-timepolymerase chain reaction experiments,βTheScientificWorld Journal, vol. 11, pp. 1383β1393, 2011.
[10] H. H. Lai, Y. C. Lin, and C. H. Yeh, βForm design of productimage using grey relational analysis and neural network mod-els,β Computers and Operations Research, vol. 32, no. 10, pp.2689β2711, 2005.
[11] B. Kim, J. Lee, J. Jang,D.Han, andK.-H.Kim, βPrediction on theseasonal behavior of hydrogen sulfide using a neural networkmodel,βTheScientificWorldJournal, vol. 11, pp. 992β1004, 2011.
[12] S. Haykin, Neural Networks: A Comprehensive Foundation,Prentice Hall, Upper Saddle River, NJ, USA, 1999.
[13] Y. Kwon, O. A. Omitaomu, and G.-N. Wang, βData miningapproaches for modeling complex electronic circuit designactivities,β Computers and Industrial Engineering, vol. 54, no. 2,pp. 229β241, 2008.
[14] R. H. Myers, Classical and Modern Regression with Application,PWSKENT, Boston, Mass, USA, 2nd edition, 1990.
[15] M. Negnevitsky, ArtiIcial Intelligence, Addison-Wesley, NewYork, NY, USA, 2002.
[16] D. McDonagh, A. Bruseberg, and C. Haslam, βVisual productevaluation: exploring usersβ emotional relationships with prod-ucts,β Applied Ergonomics, vol. 33, no. 3, pp. 231β240, 2002.
[17] N. Cross, Engineering Design Methods: Strategies for ProductDesign, John Wiley & Sons, Chichester, UK, 1994.
Submit your manuscripts athttp://www.hindawi.com
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttp://www.hindawi.com
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Journal of
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
CombinatoricsHindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
International Journal of
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
The Scientific World JournalHindawi Publishing Corporation http://www.hindawi.com Volume 2014
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttp://www.hindawi.com
Volume 2014 Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Stochastic AnalysisInternational Journal of