ANFIS modeling for predicting affective responses to tactile textures

ANFIS Modeling for Predicting Affective Responsesto Tactile TexturesDiyar Akay,1 Xiaojuan Chen,2 Cathy Barnes,2 and Brian Henson2

1 Department of Industrial Engineering, Gazi University, Ankara, Turkey2 School of Mechanical Engineering, University of Leeds, Leeds, UK

Abstract

The Adaptive Neuro-Fuzzy Inference System (ANFIS) is proposed to simulate and analyze the mappingbetween the physical properties of tactile textures and people’s affective responses. People were askedto rate the tactile feeling of 37 tactile textures against six pairs of adjectives on a semantic differentialquestionnaire. The friction coefficient, average roughness, compliance, and a thermal parameter ofeach tactile texture were measured. ANFIS models were built to predict the affective responses totactile textures. The resulting ANFIS models demonstrated a good match between predicted and actualresponses, and always yielded better performance when compared to linear and exponential regressionmodels. The effects of physical properties of textures on affective responses were also analyzed bysimulating the synthetic data with ANFIS. C© 2011 Wiley Periodicals, Inc.

Keywords: Tactile texture; Affective engineering; Feeling; Semantic; ANFIS

1. INTRODUCTIONProduct design has been highlighted as a key factor tosurvive in today’s highly competitive business environ-ment (Trueman & Jobber, 1998). The critical processbehind the success of product design is the ability tocapture and understand consumer needs thoroughly(Karkkainen, Piippo, & Tuominen, 2001). With thedevelopment of global markets and advances in pro-duction technology, where many similar products arenow often competitive in terms of price, quality, andfunctionality, satisfaction of consumers’ affective needshas emerged as an important value-added aspect in thesuccessful marketing of products (Jordan, 2000; Khalid

Correspondence to: Diyar Akay, Department of In-dustrial Engineering, Gazi University, Maltepe, Ankara06570, Turkey. Phone: +90 (0) 312 5823844; e-mail:[email protected]

Received: 6 April 2010; revised 6 October 2010; accepted 1November 2010

View this article online at wileyonlinelibrary.com/journal/hfm

DOI: 10.1002/hfm.20268

& Helander, 2006; Petiota & Yannou, 2004). An affect isdefined as a short term, discrete, conscious, nonreflec-tive, subjective feeling that may have an influence on aperson’s overall emotion (Russel, 2003). Positive affec-tive responses to products influence the decision to pur-chase a product (Holbrook & Hirschman, 1982; Khalid,2006), and thus have a positive impact on gaining mar-ket advantage. This phenomenon is particularly im-portant for well-established consumer products such ascars, mobile phones, electronic appliances, and furni-ture (Lai, Chang, & Chang, 2005). As a result, it has be-come imperative for manufacturers to understand howconsumers perceive products and how to reflect theirsubjective feelings and preferences in product design.

Affective engineering has developed to meet thisevolving trend in consumer preferences for moreaffectively engaging products. It is concerned withmeasuring people’s affective responses to products,identifying the properties of the products to whichthey are responding, and then using the informationto design better products. It is a westernized ap-proach to Kansei engineering, which was pioneered byNagamachi in Japan in the 1970s (Nagamachi, 1995).

Human Factors and Ergonomics in Manufacturing & Service Industries 22 (3) 269–281 (2012) c© 2011 Wiley Periodicals, Inc. 269

ANFIS Modeling Akay et al.

In Japanese, “kansei” means consumers’ psychologicalimpressions and feelings about a product. Kansei engi-neering has been applied successfully to many differenttypes of products (Schutte, 2005). What makes affec-tive engineering distinct from product design is that itis a process for systematically linking the consumers’feeling (affect, kansei) of a product to the product’sphysical elements (Henson, Barnes, Livesey, Childs, &Ewart, 2006). In the past, this relationship was seen asa “black box” and left to the designer’s intuitive feeling,creativity, and experience. It has been argued, how-ever, that there might be differences between the per-ceptions of designers and of customers regarding thesame product (Hsu, Chuang, & Chang, 2000). Affec-tive engineering is an approach that provides evidenceof people’s experiences of products that can be used tounderpin the decisions in product design.

In one technique of affective engineering, adjec-tive pairs that consumers use to describe the prod-uct are identified and embodied into a self-report,semantic differential questionnaire (Osgood, Suci, &Tannenbaum, 1957). A representative sample of con-sumers is then asked to rate the degree to which eachword pair describes a range of products. Last, corre-lations are established between the ratings and designelements of the product. While different techniquesare available, the aim is to identify objective relation-ships between people’s affective responses and productcharacteristics.

Although multivariate regression has been generallyused to unfold the relationships between a product’sphysical elements and affective responses (Chuang &Ma, 2001; Dore, Pailhes, Fischer, & Nadeau, 2007; Han,2003; Han, Kim, Yun, Hong, & Kim, 2004; Hsu et al.,2000; Tanoue, Ishizaka, & Nagamachi, 1997), it may beinadequate to capture relationships because the ap-proach itself has an assumption of linearity, whichmake it inappropriate to analyze affective engineer-ing data which include subjectivity and tend to con-tain large noise. This problem was first discussed byShimizu and Jindo (1995), who underlined that non-linearity of affective responses could not be detectedusing multiple regression. This nonlinearity also hasbeen acknowledged by researchers in affective engi-neering (Hotto & Hagiwara, 2006; Nagamachi, 2006;Onisawa, 2000). In this regard, alternative methodsmust be used to ensure the relationships.

Currently, soft computing is probably the mostappropriate way to identify nonlinear relationshipsbetween design elements of a product and affective

responses. In the last five years or so, the use of softcomputing techniques to map affective responses todesign elements in affective engineering has emergedas a substantial research area. Soft computing is a col-lection of techniques (fuzzy logic, neurocomputing,evolutionary computing, probabilistic computing), theaim of which is to exploit the tolerance for impre-cision and uncertainty to achieve tractable, robust,and low-cost solutions (Zadeh, 1994). There have beenmany successful examples of the use of soft computingtechniques in affective engineering using fuzzy rule–based models (Akay & Kurt, 2009; Hotto & Hagiwara,2006; Lin, Lai, & Yeh, 2007; Park & Han, 2004), roughsets (Nagamachi, Okazaki, & Ishikawa, 2006; Nishino,Nagamachi, & Sakawa, 2006; Yanagisawa & Fukuda,2005; Zhai, Khoo, & Zhong, 2009), neural networks(Chen, Khoo, & Yan, 2006; Hsiao & Huang, 2002;Ishihara, Ishihara, Nagamachi, & Matsubara, 1997; Lai,Lin, & Yeh, 2005; Lai, Lin, Yeh, & Wei, 2006), and asso-ciation rule mining (Jiao, Zhang, & Helander, 2006).

When consumers interact with products, all senses(visual, auditory, olfactory, gustatory, and tactile) havethe potential to influence affect. Most studies, however,have been concerned with the affective influences ofsenses other than the tactile (Choia & Jun, 2007). It hasbeen recognized that tactile affect plays an importantrole in affecting the attitudes of consumers (Barnes,Childs, Henson, & Southee, 2004; Grohmanna,Spangenberg, & Sprott, 2007). While there has been ex-tensive research into the perception of touch (Hollins& Bensmaiea, 2007), there has not been much workon the affective influence of touch on products suchas mobile phones, car interiors, and sports equipment.Yun, You, Geum, and Kong (2004) studied the tactilefeeling of materials for vehicle interiors (i.e., crash pad,steering wheel, wood grain, and metal grain). Barneset al. (2004) investigated how surface roughness af-fects a person’s feeling when touching glass surfaces forcosmetics’ packaging. Bahn, Lee, Lee, and Yun (2007)studied both the look and feel of car crash pads andidentified softness of the surface as the most impor-tant parameter. Childs and Henson (2007) conductedboth self-report and contact mechanism experimentsto explore people’s feelings and impressions on touch-ing screen-printed surfaces and the physical features ofa surface to which they are responding. Liu, Yue, Cai,Chetwynd, and Smith (2008) analyzed the correlationsbetween measured surface properties and perceivedfeelings for car passenger vehicle interior components.Karana, Hekkert, and Kandachar (2009) investigated

270 Human Factors and Ergonomics in Manufacturing & Service Industries DOI: 10.1002/hfm

Akay et al. ANFIS Modeling

the meaning of material through sensorial properties(roughness, glossiness, transparency, etc.) and manu-facturing processes (polishing, joining, molding, etc.).Thus, although some affective engineering research hasbeen conducted on affective aspects of touch, there isstill much work needed to identify how tactile affectrelates to the physical parameters of surface materialsof different products.

The aim of this study is to simulate and analyzethe relationships between the physical properties oftactile textures and people’s affective responses usingan Adaptive Neuro-Fuzzy Inference System (ANFIS).ANFIS, a soft computing tool, is a fuzzy inference sys-tem (FIS) implemented in a framework of adaptiveneural networks (Jang, 1993; Jang, Sun, & Mizutani,1997). In this study, ANFIS models are developed forpredicting affective responses to tactile texture as afunction on friction coefficient (μ), average roughness(Ra), rate of cooling on touching (dT/dt), and com-pliance (c). Section 2 explains the theory of ANFIS.Sections 3.1 and 3.2 describe the methods of collectionof people’s affective responses to tactile textures andphysical measurements of tactile textures, respectively.Section 3.3 explains how the ANFIS model is imple-mented. Results are presented in Section 4, followed bythe conclusion in Section 5.

2. THEORY2.1. ANFIS

A FIS implements a nonlinear mapping from an inputspace to an output space by a number of fuzzy rulesconstructed from human knowledge. The success of aFIS depends on the identification of the fuzzy rules andmembership functions tuned to a particular applica-tion. It is usually difficult in terms of time and cost, andsometimes impossible, however, to transform humanknowledge into a rule base (Nauck, Klawonn, & Kruse,1997), and, even if a rule base is provided, there remainsa need to tune the membership functions to enhancethe performance of the mapping. Neuro-fuzzy systemsovercome these limitations by using artificial neuralnetworks to identify fuzzy rules and tune the parame-ters of membership functions in FIS automatically. Inthis way, the need for the expert knowledge usually re-quired to design a standard FIS is eliminated. There areseveral approaches to integrate ANNs and FISs depend-ing on the application type (Nauck et al., 1997). A spe-cific approach in neuro-fuzzy systems is ANFIS, which

is a Sugeno type FIS implemented in the framework ofadaptive neural networks (Jang, 1993). Although it wasone of the first hybrid neuro-fuzzy models, it remainsthe best function approximator among the neuro-fuzzymodels.

To present the ANFIS architecture, a FIS with two in-puts (x1 and x2) and one output (y) is considered whereeach input is assumed to have two fuzzy sets (Kaya,Hasiloglu, Bayramoglu, Yesilyurt, & Ozok, 2003; Wang,& Elhag, 2008). The associated first-order Sugeno typefuzzy rules are then as follows:

Rule 1: If (x1 is A1) and (x2 is B1) then f11 = p11x1 +q11x2 + r11




In Sugeno fuzzy rules, the consequent parts of the rulesare the linear combination of input signals plus a con-stant term. These parameters, pij, qij, and rij, are deter-mined during the training phase of ANFIS.

The ANFIS architecture is depicted in Figure 1. Ithas five layers, and nodes in each layer have differentfunctionality. A circle indicates a fixed node, whereasa square indicates an adaptive node the parameters ofwhich are changed during the training process.

Each node in the first layer is adaptive and generatesa membership grade for each input variable. The nodeoutput (o) in this layer is defined as:

O1Ai

= μAi (x1), i = 1, 2 [1]

O1Bj

= μBj (x2), j = 1, 2 [2]

where μAi (and μBj ) are fuzzy membership functionsdefined for fuzzy levels Ai (and Bj) for the correspond-ing input x1 (and x2). For example, if the generalizedbell membership function is employed, μAi(x1) is givenby

μAi (x1) = 1/

[1 +

∣∣∣∣x1 − ci

ai

∣∣∣∣2bi

], [3]

where ai, bi, and ci are the parameters of the member-ship functions that are optimized during the trainingstage. Other continuous fuzzy membership functionscan also be used to define fuzzy levels.

In the second layer, every node is a fixed node inwhich the node function represents the firing strength

Human Factors and Ergonomics in Manufacturing & Service Industries DOI: 10.1002/hfm 271


Figure 1 ANFIS architecture.

of a rule and is computed as

O2i,j = Wi,j = μAi (x1)μBj

(x2), i, j = 1, 2. [4]

In layer 3, each node is also a fixed node and is the ratioof the rule’s firing strength to the sum of all rules’ firingstrengths.

O3i,j = Wi,j = Wi,j

W1,1 + W1,2 + W2,1 + W2,2,

i, j = 1, 2 [5]

The output of each node in the fourth layer is theproduct of the normalized firing strength found in theprevious step and the first-order polynomial of the rule

O4i,j = Wi,j

∗f i,j = Wi,j(pijx1 + qijx2 + rij),

i, j = 1, 2. [6]

In the fifth layer, a single node computes the overalloutput as the summation of all signals from Layer 4.

O5 =2∑

i=1

2∑j=1

Wi,jfi,j [7]

In the ANFIS architecture, there are two adaptive lay-ers: Layer 1 and Layer 4. Adjustable parameters in Layer1 describe the shape of the membership functions andare referred to as premise parameters. The adjustableparameters in Layer 4 related to the first-order poly-nomials are called consequent parameters. The task ofANFIS in learning is to tune the premise and conse-quent parameters until the desired input–output map-ping from the FIS is achieved. This learning task isaccomplished by a hybrid algorithm combining theleast squares method and the gradient descent method

(Jang, 1993). The hybrid algorithm is composed ofa forward pass and a backward pass. In the forwardpass of the algorithm, the premise parameters are heldfixed, and functional signals go forward to Layer 4;then the consequent parameters are determined by theleast squares method. In the backward pass, conse-quent parameters are held fixed, and the error measurepropagates backward; then the premise parameters areupdated by the gradient descent method to adjust themembership functions.

2.2. Testing, Validation, andPerformance of ANFIS Models

K-fold cross-validation (Stone, 1974) can be used totrain and validate ANFIS models and is the approachtaken here. This approach is suitable for a small data set,has the advantage that data are used for both trainingand testing, and avoids overfitting. Each response to thestimulus in the data is used to test the model exactlyonce, and is included in training set k-1 times.

To assess the performance of ANFIS models, meanabsolute percent error (MAPE), and correlation coef-ficient (R) criteria are used in this research. They aredefined as follows:

MAPE = 1

n

n∑i=1

|REi| × 100 and [8]

REi =∣∣∣∣ai − pi

ai

∣∣∣∣ , [9]

where ai and pi are actual and predicted values, respec-tively. RE is absolute relative error, and n is the numberof training or testing samples. The smaller the MAPE



values, the better the performance. A value of MAPEof less than 10% is regarded as excellent (Chang, Liu,& Lai, 2008). MAPE is used to determine when to stoptraining. The error measure of the test set decreasesduring the initial epochs of the training phase. Trainingis stopped when the MAPE error for test set starts to in-crease, indicating that a minimum has been achieved.The optimal premise and consequent parameters ofeach FIS are tuned for the model that has the structureparameters that provide the minimum number of testerrors.

R is estimated to measure how well the variation inthe predicted outputs is explained by observed values.It is defined as

R =∑N

i=1 (ai − a)(pi − p)√∑Ni=1 (ai − a)2

√∑Ni=1 (pi − p)2

, [10]

where a and p are the mean values of the actualand predicted data sets, respectively. It shows the de-gree of association between predicted and actual val-ues. The model prediction is good if the value of R

approaches 1.The threshold statistic (TS) is used to provide infor-

mation on the distribution of errors. It is computed fora x% level as

TSx = Yx

n100, [11]

where Yx is the number of predicted responses forwhich the RE is less then x%, and n is the total numberof predicted responses over test partitions of data withANFIS.

3. APPLICATION OF ANFIS TO THEAFFECTIVE DESIGN OF TACTILETEXTURES3.1. Measurement of AffectiveResponses

Eighteen participants (12 men and 6 women), from20 to 60 years old, and who responded to postersand e-mails inviting volunteers, participated in the ex-periment. Thirty-seven materials with different tex-tures were used as stimuli (Figure 2). The stimulican be grouped into three categories: 1) Stimuli 1–22were cardboards; 2) 23–31 were papers and foils; and3) stimuli 32–37 were laminate boards. The cardboards,papers, and foils were packaging materials used for con-fectionery items. The laminate boards were samples of

Figure 2 Stimuli.

materials typically used for making office furniture andwere included to increase the variety of textures. As aresult, the 37 stimuli covered a variety of textures withdifferent physical properties, such as roughness, hard-ness, and thermal conductivity. The stimuli were cutinto 10 cm × 8 cm rectangles and were numbered foridentification.

The tactile properties of the stimuli were rated byparticipants against six pairs of adjectives. The adjec-tive pairs were “slippery–sticky,” “bumpy–flat,” “wet–dry,” “hard–soft,” “smooth–rough,” and “warm–cold.”These adjectives were used because they are often usedin affective engineering experiments concerning touchin a variety of contexts. Before the experiment, the pur-pose and procedures of the experiment were explainedto the participants. The participants were informed ofthe definition of each adjective pair and carried outpractice trials to become familiar with the experimen-tal task. The stimuli were presented to the participantsin boxes so that they could not be seen. One side ofeach box was kept open and was covered with a smallwhite curtain to prevent sight of the stimuli while stillallowing participants to touch them. Each participantwas asked to place his or her hands into the box underthe white curtain and touch one texture at a time. Thestimuli were presented in a random order. Participantswere allowed to use active dynamic touch to explorethe surface of each texture. No restrictions were givenas to which hand or parts of the hand could be used orfor how long the stimuli could be inspected. After feel-ing each stimulus, the participants were asked to ratethe texture against each adjective pair on a 20-pointscale (Figure 3). The adjective pairs were presentedin a random order and in random polarity on each



warm cold

Figure 3 An example scale used in the experiment.

questionnaire. After each stimulus had been evaluated,it was removed from the box. Participants received asmall gift at the end of the session to compensate themfor their time.

3.2. Measurement of PhysicalProperties

Although there are many computational texture fea-tures (Hasiloglu, 2001), there are a few physical mea-surements related to tactile feeling. In this study, thefriction coefficient (μ), average roughness (Ra ; a ther-mal property), rate of cooling on touching (dT/dt),and compliance (c) of the 37 tactile textures were mea-sured. Arithmetical mean roughness, Ra (μm), wasused and is the most popular measure of roughness oftextures. It is the integral of the deviations from themean of the heights of the peaks and valleys of thesurface. Roughness was measured using a commercialstylus surface profilometer (RTH Form Talysurf 120L;Taylor-Hobson, Leicester, UK).

Values of μ, dT/dt, and c were measured on apiezoelectric force platform (Kistler MiniDyn, Ost-fildern, Germany) (Figure 4). For the friction measure-ment, each stimulus was fixed to the force platform(Figure 4a). An experimenter pressed (load Fy) andslid (load Fx) her fingertip against it. Fx and Fy wererecorded against time. The friction coefficient was ob-

X

Y

Sample

x, y force platform

(b) (c)Time

FX, FY (N)

FY

FX

Time

Finger movement TC

T (°C)

dY Steel ball

Soft support

FY (N)

dY (mm)

Figure 4 Schematic views of a) friction, b) heat transfer, c) compliance measurements.

tained from Fx/Fy. Loads Fy were in the range 0.5to 3N.

To measure the rate of cooling of the finger whenthe stimulus is touched, an artificial, silicone rubberfingertip was loaded on to the surface, without sliding(FY = 1 N) (Figure 4b). A thermocouple (TC) was em-bedded just within the tip. Before contacting the sur-face, the tip was heated by an internal cartridge heaterto the temperature of the skin of the human finger,32 ± 0.2◦C. On contact between the loaded finger andthe stimulus, the decrease in the temperature recordedby the TC against time was recorded. The maximumrate of change (◦Cs−1), which occurred at the start ofcontact, was taken as the measure of the rate of cooling.

The value of c was measured in the following way(Figure 4c): A soft rubber support was inserted be-tween the stimulus and the force platform. The stimu-lus was pressed with a steel ball with a radius of 7 mm.With a 4N loading force, the ball’s displacement withincreasing load was recorded. The measure of c wasempirically taken to be the value of the displacement(mm) when Fy = 3N. In all cases, measurements wererepeated several times and averages obtained.

3.3. Implementation of the ANFISModels

The aim of the ANFIS was to model and predict theaffective response to texture as a function of μ, Ra ,dT/dt, and c. To do this, six separate ANFIS modelswere built, one for each of the six adjective pairs. Ineach ANFIS model, μ, Ra , dT/dt, and c were takenas the input, and the mean average response to a



Figure 5 ANFIS structure of the model.

particular adjective pair was used as the output. TheANFIS structure used in this application is shown inFigure 5, where two fuzzy levels are assigned to eachinput (i.e., small and large linguistic labels). The gen-eralized bell membership function was used to definefuzzy levels because of its smoothness and concisenotation. The total number of rules in the structureis 16.

The ANFIS models were implemented using MAT-LAB (version 7.3; MathWorks, Natick, MA) with theFuzzy Logic Toolbox. A hybrid learning algorithm wasused. Structure parameters, such as number of epochsand learning rate, were determined by trial and errorby carrying out preliminary runs.

The data for one stimulus were removed during pre-liminary training runs of ANFIS because they had anegative effect on the performance of the models. Theremoved sample had the second highestRa value (13.02μm), was above average in roughness (2.25 μm), andwas considered as an outlier. The removed sample feltqualitatively different from the other stimuli with anincongruent rough, metallic feel. The remaining 36stimuli were used for ANFIS modeling.

K-fold cross-validation (Stone, 1974) was used totrain and validate the ANFIS model. The remaining 36stimuli were randomly divided into nine subsets, eachof which contained four cases. The nine ANFIS modelswere trained, each time one of the nine subsets was as-

signed to a test set and the other eight subsets were allo-cated to a training set. The overall performance was cal-culated as the average performance of the nine modelsover the corresponding test partitions of the data. Theperformance of the models was assessed using MAPEand Pearson’s product moment correlation coefficient,R. To compare the performance of the ANFIS mod-els, linear and exponential regression equations werealso built. Exponential equations were in the form ofresponse = b0(μb1 )(Rb2

a )(dT/dtb3 )(cb4 ), where b0, b1,b2, b3, and b4 were the estimated regression coeffi-cients. By applying logarithmic transformations, non-linear exponential forms were converted into linearform, and then regression analysis using least squareestimation was applied to compute the coefficients ofthe exponential model. Ninefold cross-validation wasagain applied to build and test the regression models.The test sets used in regression models were the sameas those already generated for the ANFIS models.

Because even poor models can produce high corre-lations (80 or 90%), since correlation only indicatesthe strength of the linear relationship between the ob-served and the predicted values, the TS was calculatedfor different levels of absolute error.

To show the effects of physical properties on affec-tive responses, synthetic data were generated using 34

full factorial design with three levels for each physicalcharacteristic, comprising a total of 81 stimuli. Three



TABLE 1. Physical Characteristics and Mean Scores of Adjectives for 37 Tactile Textures

Physical Characteristics Affective Responses

dT/dt c Slippery Bumpy– Wet– Hard– Smooth– Warm–Stimuli μ Ra (μm) (◦Cs−1) (mm) –Sticky Flat Dry Soft Rough Cold

1 0.29 0.74 1.88 0.58 6.06 17.44 15.44 8.33 4.06 11.062 0.26 13.02 2.07 0.60 11.39 3.17 14.06 7.44 16.50 11.393 0.23 3.10 1.60 0.65 7.44 7.22 17.44 10.78 14.44 8.22

· · · · · · · · · · ·13 0.57 0.07 2.12 0.64 9.18 18.50 10.61 8.22 3.06 10.06· · · · · · · · · · ·36 0.26 3.78 2.90 0.44 10.72 9.67 14.78 4.44 11.22 13.5037 0.53 0.05 3.17 0.47 7.39 19.50 10.22 4.72 1.94 14.94

The data from stimulus 2 were not used.

levels were defined by taking minimum, average, andmaximum values of physical characteristics of origi-nal stimuli. The data were then simulated with ANFIS,and the results were plotted where the point in the plotsrepresents the mean predicted response at each level ofthe physical characteristic while changing the levels ofother characteristics.

4. RESULTS AND DISCUSSIONTable 1 presents mean averages of the participants’responses to each stimulus against the six adjectivepairs and the stimuli’s physical properties. The val-ues of the friction coefficient, μ, for the stimuliranged between 0.17 and 0.84 and averaged 0.32. Thevalues for roughness, Ra ranged from 0.07 μm to13 μm and averaged 2.25 μm. The value of the cool-ing rate, dT/dt, ranged from 0.35◦Cs−1 to 3.29◦Cs−1

and averaged 1.88◦Cs−1. The value of the compliance,c, ranged from 0.41 mm to 1.18 mm and averaged0.72 mm.

The training and test performances of the optimizedANFIS models are presented in Table 2. MAPE valuesin Table 2 were the average result achieved via ninefoldcross-validation. Satisfactory results were achieved forthe ANFIS models of “wet–dry,” “hard–soft,” “warm–cold,” and “slippery–sticky” adjective pairs. MAPE val-ues of these four models were lower than 10%, indi-cating that the relationship between physical charac-teristics of textures and each of these four adjectivepairs is highly significant. No strong relationship wasobserved for the ANFIS models of “bumpy–flat” or“smooth–rough” adjective pairs.

The MAPE values indicate that the ANFIS mod-els always yielded lower errors when compared to re-gression models. High error values were observed for

TABLE 2. Training and Test Performance of ANFIS Models

ANFIS∗Regression (Linear)∗ Regression (Exponential)∗

MAPE MAPE MAPE MAPEModel for (training) (test) (test) (test)

Slippery–Sticky 0.09 10.05 15.00 15.51Bumpy–Flat 0.28 19.83 31.25 33.48Wet–Dry 0.03 3.61 6.85 6.88Hard–Soft 0.07 5.80 11.16 12.85Smooth–Rough 0.39 35.50 46.07 35.76Warm–Cold 0.09 6.40 11.85 10.50

∗Ninefold cross-validation is applied.



Figure 6 Scatter diagrams showing actual and predicted responses (for test set).

“bumpy–flat” and “smooth–rough” adjective pairs inboth the linear and exponential regression models. Al-though the performance of the exponential model for“smooth–rough” is close to that of the ANFIS model,it is too high to be significant.

Figure 6 shows the predicted outputs of the sixANFIS models against their corresponding actual re-sponses. It was found that best fit lines (red lines) weresituated close to unit lines (dotted lines). High corre-lations between actual and predicted values were ob-served for the models for “slippery–sticky,” “wet–dry,”“hard–soft,” and “warm–cold.” This shows that pre-dicted responses of the ANFIS models have a goodmatch with actual responses.

Figure 7 presents TS distribution errors for ANFISand regression models for different threshold levels.For the “slippery–sticky” adjective pair, it was observedthat while 70% of the predicted responses are within the10% error level for the ANFIS, it is approximately 20%for regression models. For “hard–soft” and “warm–

cold” adjective pairs, it is less than 10% for ANFISas opposed to between 10% and 20% for regressionmodels. The TS is less than 10% for both ANFIS andregression models, but it is still lower in ANFIS for the“wet–dry” adjective pair. This finding demonstratesthat ANFIS has better performance than the regressionmodels.

The reason why ANFIS has not achieved satisfac-tory results for “bumpy–flat” and “smooth–rough” ad-jective pairs might be twofold. First, it might be thatstimuli could not be evaluated reliably for these twoadjective pairs by participants. It was observed thatvariances of responses to “bumpy–flat” and “smooth–rough” are higher than those of responses to otheradjective pairs. Second, some of the physical charac-teristics of the tactile texture might be irrelevant formodeling the “bumpy–flat” and “smooth–rough” ad-jective pairs.

Using synthetic data generated, main effect plots forμ, Ra , and rate dT/dt are drawn in Figure 8.



Figure 7 Distribution of errors for ANFIS and regression models.

The following conclusions are drawn fromFigure 8.

• It is pointed out that four physical propertiesare highly significant for the “slippery–sticky”response. Effects of all physical properties arepositive.

• Remarkable main effects on the response “wet–dry” are reported for Ra , dT/dt, and c. It canbe seen that effects of dT/dt and c are nega-tive, whereas the effect of Ra is positive on theresponse. There is no significant interaction be-tween μ and “wet–dry” response.

• Ra , dT/dt, and c are highly significant for the“hard–soft” response. μ seems not to have a sig-nificant effect. It can be seen that Ra has a nega-tive strong effect, whereasc has a positive strongeffect on the response, and the “hard–soft” re-sponse is more sensitive to dT/dt changes.

• dT/dt is the only significant physical propertyon the response “warm–cold.

Most of the stimuli (approximately 86%) used in thisexperiment were examples of packaging materials usedfor confectionery items. It is likely therefore that, al-though the stimuli were not presented to participants



Figure 8 Main effects plots for affective responses.

in the context of confectionery items, the results of thisexperiment can be generalised and used to predict hu-man responses to confectionery packaging materials.Whether the results can be generalized further (e.g.,

to materials with similar ranges of physical properties)cannot yet be determined, because the effect of con-text on people’s perceptions of products is not yet fullyunderstood.

The use of ANFIS is a first step in using soft comput-ing techniques for the analysis of affective responsesto tactile textures. Through its use we will be betterable to determine the effects of tactile texture prop-erties on people’s feelings about products, and, withfurther iterations, choose a more comprehensive rangeof stimulus materials that is representative of the hu-man experience of touch.

5. CONCLUSIONSIn this study, ANFIS was used to simulate and an-alyze the relationship between the tactile texturesand people’s affective responses. The structure of theANFIS, the way in which data were collected, and howthe ANFIS was modeled were described. The ANFISmodel performed well. MAPE values of the test setwere approximately 10% or better, indicating a goodmatch between predicted and actual responses, and theANFIS models always yielded lower errors when com-pared to regression models. It is proposed that this isthe principal advantage of using ANFIS in this applica-tion. ANFIS performance over regression models wasalso demonstrated using the TS parameter. The effectsof physical characteristics on affective responses werediscussed by simulating new data and then drawingmain effect plots. Such models could be used to predictpeople’s affective responses to different combinationsof properties of products and, as a result, remove theneed for consumer studies.

ACKNOWLEDGMENTSThis work was carried out while Dr. Akay was a Uni-versity of Leeds postdoctoral research fellow, fundedby The Scientific and Technological Research Councilof Turkey (TUBITAK). The work was partly funded bythe UK’s Engineering and Physical Sciences ResearchCouncil (EPSRC; EP/D060079/1) and the EuropeanCommission (EC; NEST 043157). The views expressedhere are the authors’ and not those of the EC.

References

Akay, D., & Kurt, M. (2009). A neuro-fuzzy based ap-proach to affective design. The International Journal of



Advanced Manufacturing Technology, 40(5–6), 425–437.

Bahn, S., Lee, C., Lee, J. H., & Yun, M. H. (2007). A statis-tical model of relationship between affective responsesand product design attributes for capturing user needs.Lecture Notes in Computer Science, 4560, 305–313.

Barnes, C. J., Childs, T. H. C., Henson, B., & Southee, C.H. (2004). Surface finish and touch–a case study in anew human factors tribology. Wear, 257, 740–750.

Chang, P. C., Liu, C. H., & Lai, R. K. (2008). A fuzzy case-based reasoning model for sales forecasting in printcircuit board industries. Expert Systems with Applica-tions, 34(3), 2049–2058.

Chen, C. H., Khoo, L. P., & Yan, W. (2006). An investiga-tion into affective design using sorting technique andKohonen self-organising map. Advances in EngineeringSoftware, 37, 334–349.

Childs, T. H. C., & Henson, B. (2007). Human tactileperception of screen-printed surfaces: Self-report andcontact mechanics experiments. Proceedings of the In-stitution of Mechanical Engineers, Part J: Journal ofEngineering Tribology, 221(3), 427–441.

Choia, K., & Jun, C. (2007). A systematic approach tothe kansei factors of tactile sense regarding the surfaceroughness. Applied Ergonomics, 38(1), 53–63.

Chuang, M. C, & Ma, Y. C. (2001). Expressing the ex-pected product images in product design of micro-electronic products. International Journal of IndustrialErgonomics, 27, 233–245.

Dore, R., Pailhes, J., Fischer, X., & Nadeau, J. P. (2007).Identification of sensory variables towards the inte-gration of user requirements into preliminary design.International Journal of Industrial Ergonomics, 37,1–11.

Grohmanna, B., Spangenberg, E. R., & Sprott, D. E.(2007). The influence of tactile input on the evalua-tion of retail product offerings. Journal of Retailing,83(2), 237–245.

Han, S. H. (2003). A systematic approach for couplinguser satisfaction with product design. Ergonomics,46(13–14), 1441–1461.

Han, S. H., Kim, K. J., Yun, M. H., Hong, S. W., & Kim, J.(2004). Identifying mobile phone design features criti-cal to user satisfaction. Human Factors and Ergonomicsin Manufacturing, 14(1), 15–29.

Hasiloglu, A. S. (2001). Rotation-invariant texture anal-ysis and classification by artificial neural networks andwavelet transform. Turkish Journal of Engineering andEnvironmental Sciences, 25, 405–413.

Henson, B., Barnes, C. J., Livesey, R., Childs, T. H. C.,& Ewart, K. (2006). Affective consumer requirements:A case study of moisturizer packaging. ConcurrentEngineering: Research and Applications, 14(3), 187–196.

Holbrook, M. B., & Hirschman, E. C. (1982). The expe-riential aspects of consumption: Consumer fantasies,feeling, and fun. The Journal of Consumer Research, 9,132–140.

Hollins, M., & Bensmaiea, S. J. (2007). The coding ofroughness. Canadian Journal of Experimental Psychol-ogy, 61(3),184–195.

Hotto, H., & Hagiwara, M. (2006). A fuzzy rule based per-sonal kansei modeling system. In IEEE InternationalConference on Fuzzy Systems (pp. 1031–1037). Van-couver, Canada.

Hsiao, S. W., & Huang, H. C. (2002). A neural networkbased approach for product form design. Design Stud-ies, 23, 67–84.

Hsu, S. H., Chuang, M. C., & Chang, C. C. (2000). A se-mantic differential study of designers’ and users’ prod-uct form perception. International Journal of IndustrialErgonomics, 25, 375–391.

Ishihara, S., Ishihara, K., Nagamachi, M., & Matsubara, Y.(1997). An analysis of kansei structure on shoes usingself-organizing neural networks. International Journalof Industrial Ergonomics, 19, 93–104.

Jang, J. S. R. (1993). ANFIS: Adaptive network-basedfuzzy inference system. IEEE Transactions on Systems,Man and Cybernetics, 23(3), 665–685.

Jang, J. S. R., Sun, C. T., & Mizutani, E. (1997). Neuro-fuzzy and soft computing: A computational approach tolearning and machine intelligence. Upper Saddle River,NJ: Prentice-Hall.

Jiao, J. R., Zhang, Y., & Helander, M. (2006). A kanseimining system for affective design. Expert Systems withApplications, 30(4), 658–673.

Jordan, P. W. (2000). Designing pleasurable products: Anintroduction to the new human factors. London: Taylor& Francis.

Karana, E., Hekkert, P., & Kandachar, P. (2009). Meaningsof materials through sensorial properties and manu-facturing processes. Materials & Design, 30(7), 2778–2784.

Karkkainen, H., Piippo, P., & Tuominen, M. (2001). Tentools for customer-driven product development in in-dustrial companies. International Journal of Produc-tion Economics, 69, 161–176.

Kaya, M., Hasiloglu, A. S., Bayramoglu, M., Yesilyurt, H.,& Ozok, A. F. (2003). A new approach to estimate an-thropometric measurements by adaptive neuro-fuzzyinference system. International Journal of Industrial Er-gonomics, 32, 105–114.

Khalid, H. M. (2006). Embracing diversity in user needsfor affective design. Applied Ergonomics, 37(4), 409–418.

Khalid, H. M., & Helander, M. G. (2006). Customer emo-tional needs in product design. Concurrent Engineer-ing: Research and Applications, 14(3), 197–206.



Lai, H. H., Chang, Y. M., & Chang, H. C. (2005). A robustdesign approach for enhancing the feeling quality of aproduct: A car profile case study. International Journalof Industrial Ergonomics, 35(5), 445–460.

Lai, H. H., Lin, Y. C., & Yeh, C. H. (2005). Form design ofproduct image using grey relational analysis and neuralnetwork models. Computers and Operations Research,32, 2689–2711.

Lai, H. H., Lin, Y. C., Yeh, C. H., & Wei, C. H. (2006).User-oriented design for the optimal combination onproduct design. International Journal of ProductionEconomics, 100(2), 253–267.

Lin, Y. C., Lai, H. H., & Yeh, C. H. (2007). Consumer-oriented product form design based on fuzzy logic: Acase study of mobile phones. International Journal ofIndustrial Ergonomics, 37, 531–543.

Liu, X., Yue, Z., Cai, Z., Chetwynd, D. G., & Smith, S. T.(2008). Quantifying touch–feel perception: Tribologi-cal aspects. Measurement Science and Technology, 19,084007.

Nagamachi, M. (1995). Kansei engineering: A new er-gonomic consumer-oriented technology for productdevelopment. International Journal of Industrial Er-gonomics, 19, 93–104.

Nagamachi, M. (2006). Kansei engineering and roughsets model. Lecture Notes in Computer Science, 4259,27–37.

Nagamachi, M., Okazaki, Y., & Ishikawa, M. (2006). Kan-sei engineering and application of the rough sets model.Proceedings of the Institution of Mechanical Engineers,Part I: Journal of Systems and Control Engineering, 220,763–768.

Nauck, D., Klawonn, F., & Kruse, R. (1997). Foundationsof neuro fuzzy systems. New York: Wiley.

Nishino, T., Nagamachi, M., & Sakawa, M. (2006). Ac-quisition of kansei decision rules of coffee flavor usingrough set method. Kansei Engineering International,5(4), 41–50.

Onisawa, T. (2000). Soft computing technique in kan-sei (emotional) information processing. In Z. Q. Liu& S. Miyamoto (Eds.), Soft computing and human-centered machines (pp. 215–248). Tokyo: Springer-Verlag.

Osgood, C. E., Suci, G. J., & Tannenbaum, P. H. (1957).The measurement of meaning. Champaign, IL: Univer-sity of Illinois Press.

Park, J., & Han, S. H. (2004). A fuzzy rules based ap-proach to modelling affective user satisfaction towardsoffice chair design. International Journal of IndustrialErgonomics, 34, 31–47.

Petiota, J. F., & Yannou, B. (2004). Measuring consumerperceptions for a better comprehension, specificationand assessment of product semantics. InternationalJournal of Industrial Ergonomics, 33, 507–525.

Russel, J. A. (2003). Core affect and the psychologicalconstruction of emotion. Psychological Review, 110(1),145–172.

Schutte, S. (2005). Engineering emotional values in prod-uct design–Kansei engineering in development. Un-published doctoral dissertation, Linkoping University,Linkoping, Sweden.

Shimizu, Y., & Jindo, T. (1995). A fuzzy logic analysismethod for evaluating human sensitivities. Interna-tional Journal of Industrial Ergonomics, 15, 39–47.

Stone, M. (1974). Cross-validatory choice and assessmentof statistical predictions. Journal of the Royal StatisticalSociety, Series B (Methodological), 36(1), 111–147.

Tanoue, C., Ishizaka, K., & Nagamachi, M. (1997). Kanseiengineering: A study on perception of vehicle interiorimage. International Journal of Industrial Ergonomics,19,115–128.

Trueman, M., & Jobber, D. (1998). Compete through de-sign. Long Range Planning, 31, 594–605.

Wang, Y. M., & Elhag, T. M. S. (2008). An adaptive neuro-fuzzy inference system for bridge risk assessment. Ex-pert Systems with Applications, 34(4), 3099–3106.

Yanagisawa, H., & Fukuda, S. (2005). Interactive reductevolutional computation for aesthetic design. Journal ofComputing and Information Science in Engineering, 5,1–7.

Yun, M. H., You, H., Geum, W., & Kong, D. (2004).Affective evaluation of vehicle interior craftsmanship:Systematic checklists for touch/feel quality of surface-covering material. In Proceedings of the Human Fac-tors and Ergonomics Society 48th Annual Meeting (pp.971–975), New Orleans, Louisiana.

Zadeh, L. A. (1994). Fuzzy logic, neural networks, and softcomputing. Communications of the ACM, 37, 77–84.

Zhai, L. Y., Khoo, L. P., & Zhong, Z. W. (2009). Adominance-based rough set approach to Kansei engi-neering in product development. Expert Systems withApplications, 36, 393–402.


ANFIS modeling for predicting affective responses to tactile textures

Documents