This action might not be possible to undo. Are you sure you want to continue?

9

**Experimental Design and Response-Surface Methodology
**

Shuryo Nakai, Eunice C. Y. Li-Chan, and Jinglie Dou

CONTENTS 9.1 Introduction .......................................................................................................................... ..294 9.2 Basic Principles ....................................................................................................................294 9.2.1 Factor Screening........................................................................................................294 9.2.2 Errors ........................................................................................................................295 9.2.3 Replication ................................................................................................................295 9.3 Factorial Designs ..................................................................................................................295 9.3.1 Elements of Factorial Analysis ................................................................................295 9.3.1.1 Effects of Main Factors ..............................................................................295 9.3.1.2 Effects of Factor Interactions......................................................................296 9.3.1.3 Confounding................................................................................................296 9.3.1.4 Symmetry ....................................................................................................296 9.3.1.5 Number of Experiments..............................................................................296 9.3.2 Fractional Factorial Designs ....................................................................................296 9.3.2.1 Two-Level Design Example ......................................................................297 9.3.2.2 Three-Level Design Example ....................................................................300 9.3.2.3 How to Select Designs................................................................................306 9.4 Response-Surface Methodology ..........................................................................................308 9.4.1 Introduction................................................................................................................308 9.4.2 Central Composite Designs ......................................................................................309 9.4.3 Mixture Designs ........................................................................................................310 9.4.4 Sequential Simplex Optimization ............................................................................312 9.4.5 Random-Centroid Optimization ................................................................................314 9.4.6 Extended Application of RCO ..................................................................................316 9.4.6.1 Mixture Designs ..........................................................................................316 9.4.6.2 Food Formulation........................................................................................316 9.4.6.3 Shape Design and Market Survey ..............................................................316 9.4.6.4 Site-Directed Mutagenesis ..........................................................................317

293

q 2006 by Taylor & Francis Group, LLC

294

HANDBOOK OF FOOD AND BIOPROCESS MODELING TECHNIQUES

9.4.7 Multivariate Response Surfaces ................................................................................319 9.4.8 SSO vs. RCO ............................................................................................................319 Glossary ........................................................................................................................................319 References......................................................................................................................................320

9.1

INTRODUCTION

The basic principle of experimental design is to plan experiments so as to be able to study the effects of certain factors on some speciﬁc results, and to identify the important inﬂuential factors with a given level of conﬁdence. The effects of the factors and their interactions are evaluated by computing F-values, which are the ratio of effect-mean-square vs. error-mean-square. Randomization and minimization of error are two important mathematical principles for efﬁcient selection of truly inﬂuential factors. Randomization is useful in objectively detecting uncontrollable errors, thereby avoiding the inclusion of errors arising from man-made causes in the true errors. Sometimes “blocking” is required during randomization when there is an unavoidable need for handling many factors in an experimental plan. Blocking should be made based on a controllable condition (factor), such as date, so that the inﬂuence of that factor can be determined later if required. Response-surface methodology (RSM) is a modeling technique to illustrate the effects of factors on response surfaces for the purpose of locating the optimum. Therefore, the RSM consists of the techniques for experimental design, regression analysis, and computation of the optimum. There are two popular books for RSM: (1) Response Surface Methodology: Process and Product Optimization Using Designed Experiments by Myers and Montgomery1 and (2) Response Surfaces: Designs and Analysis by Khuri and Cornell.2 The former has a broader coverage, including the category of evolutionary operation such as sequential simplex optimization (SSO), which is not covered by the latter. To avoid conﬂict with other chapters in this handbook on linear programming and optimization, the discussion in this chapter will be restricted to the response-surface modeling, including some optimization techniques based on response surfaces. This is indispensable because the RSM itself is an optimization technology, as discussed in the above textbook by Myers and Montgomery.1 Another point to be stressed is that this chapter does not give any detailed or deep elaboration on the methods, theories, algorithms, and principles. The authors have followed the approach used in the excellent chapter entitled “Optimization and Experimental Designs” by Otto,3 which was written for a book for analytical chemists. The readers are advised to refer to his chapter for more information, as we have avoided excessive duplication and instead have focused this chapter to discuss the speciﬁc nature of techniques useful in food- and bio-processing.

9.2 9.2.1 Factor Screening

BASIC PRINCIPLES

There are two conﬂicting considerations in selecting factors for designing experiments. (1) Theoretically, all potential factors should be entered into an experimental design to avoid missing important factors that may appear minor but could in fact be critical in the mechanism of the reaction in question. (2) However, it is generally accepted that the greater the number of factors, the more the number of iterations that are required to reach the optimum. This increase in the number of experiments—usually exponential—would immediately increase labor and expenses of carrying out experiments, especially in the case of biological exercises.

q 2006 by Taylor & Francis Group, LLC

EXPERIMENTAL DESIGN AND RESPONSE-SURFACE METHODOLOGY

295

To circumvent these problems, a preliminary selection of factors is customarily made using a factorial design, especially fractional factorial design. The L16(215) design of Taguchi4 has been one of the most useful designs for this purpose due to easy identiﬁcation of factor-interaction columns in the design, thus being efﬁcient in selecting inﬂuential factors or interactions without losing important information. The delicate nature and complexity of biological phenomena can be exempliﬁed in cytokine chemistry and genetic engineering. For instance, interleukins 4, 5, and 6 may simultaneously exert their functions either antagonistically or symbiotically in the probiotic activity of orally administered lactobacilli.5 The function of single residues in a protein sequence cannot be accurately deﬁned by replacement or deletion of the site in the sequence.6 In the case of multifunctional phenomena, it is possible that the scree-plot of principal components3 does not show a sharp decline of eigenvalues, thereby implying that many factors are playing almost equally important roles in some biological phenomena. In this case, ignoring higher-order interactions for the sake of dimensionality reduction may not always be justiﬁed as in the case of Taguchi designs.4 Decision making for the priority on whether full information should be kept so as not to miss the chance of important discovery or the best efﬁcacy in experiments by sacriﬁcing some rarely signiﬁcant high level interactions is extremely critical. 9.2.2 Errors

Error is deﬁned as the cumulative effects of uncontrollable factors that are the effects of undeﬁned factors as well as any other factors that have not been selected during the designing of experiments. However, the most important error is the one that would critically affect the reliability of analysis. It is highly recommended to maintain factors as constant as possible, because the repeatability of analysis is extremely critical. Training of analysts prior to commencing experiments is a minimum prerequisite, as increasing the cycles of replication is not as efﬁcient as intensive training in terms of reducing the size of error, which directly affects the reliability of selecting the truly inﬂuential factors. 9.2.3 Replication

Replication is a popular way to decrease the size of error. However, because the error is only pﬃﬃﬃ reduced by a factor of 1= n, where n is the order of replication such that for duplication nZ2 or for triplication nZ3, replication is not a very efﬁcient way to reduce error. Therefore, for the sake of the best efﬁciency in biological projects, it is recommended to include replicates only for the purpose of computing the standard error value, instead of replicating the entire design.

9.3 9.3.1

FACTORIAL DESIGNS

Elements of Factorial Analysis

In contrast to one-factor-at-a-time approaches, where only one factor is varied while other variables are held constant, designed experiments involving factorial analysis allow the analyst to investigate the effects of the individual (main) factors as well as interactions between the factors. 9.3.1.1 Effects of Main Factors Signiﬁcance is assessed using F-values (mean square of treatment/mean square of error), as previously discussed.

q 2006 by Taylor & Francis Group, LLC

3. Consequently.2 Fractional Factorial Designs It is reasonable to believe that the higher the order of factor-factor interactions.5 Number of Experiments In full factorial designs. it is of paramount importance to obtain the desired information by most efﬁciently conducting the minimum number of experiments. In other words. all possible interactions between factors are assessed. These methods are called fractional factorial designs and are frequently used as a preliminary step during an optimization project by selecting the truly inﬂuential factors prior to performing the optimization experiments.g. and 128 experiments. whereas a three-level. respectively.1. 5-factor design requires 243 experiments. 9. If the main effects are parallel. 6. thereby analyzing main effects and lower(e. and it also increases as the number of levels of factors increases.3 Confounding Confounding. For instance. the full factorial designs for two-level experiments with 5. the less the statistical signiﬁcance. can occur between factors and/or interactions during the course of factorial analysis.1.4 Symmetry Experiments should be designed to distribute evenly within the search space of individual factors as much as possible to avoid undesirable confounding. and that unexpectedly high or low response values at speciﬁc combination of levels of more than two factors are not as important as the main effects or second-order interactions. the total number of required experiments in the design is reduced by half. This rule was utilized in the regulated random design of the random-centroid optimization (RCO) as described later. Because it is usually not only expensive but also time consuming to carry out each experiment.. second-) order interactions only. and 7 factors would require 32. if the response in question is processing yield.1. By ignoring one such interaction. especially in the life sciences. 9. Existence of signiﬁcant interaction between factors A and B means that a speciﬁc combination of factors A and B would result in an effect that is different from the unilateral effects of factors A and B considered separately. the required number of experiments increases exponentially as the number of factors (n) increases. which cannot be anticipated from knowing the main effects of A and B independently. There are such techniques available. 64.3. 9. Note that these assumptions lead to confounding of the higher-order interactions with some main effects or lower-order interactions. 9.1.296 HANDBOOK OF FOOD AND BIOPROCESS MODELING TECHNIQUES 9. which gain their efﬁciency by disregarding higher-order interactions. in which effects are inseparable.3. it is assumed that usually two factor interactions are adequate for judging the general trend in the response surface.3. For example. yet satisfactorily adequate information can be extracted.3. they are unilateral.and three-level full factorial designs require 2n and 3n experiments for n-factor analysis.2 Effects of Factor Interactions Signiﬁcant interaction between factors is deﬁned as the departure from parallelism. LLC . Two. q 2006 by Taylor & Francis Group. This is the basis of fractional factorial designs. in that only a portion of the full factorial design is chosen. a signiﬁcant factor interaction would mean that a speciﬁc combination of A and B could result in unexpectedly high increase or decrease in the yield. and all combinations of levels are taken into consideration.

the most useful L16(215) design will be explained in this chapter to illustrate the basic principles. For example. the most logical half-design is to eliminate the experimental condition appearing in column 15 with highest order of interaction. and 10. with one left-over column 15 assigned to the error term. 4. Columns 1–15 describe the experimental conditions as combinations of the two levels of factors for each of the 16 experiments in rows 1–16. and (3) ﬁnd the column that matches the computed component products. and 15) are designed to be assessed for their statistical signiﬁcance. The interaction effect of two factors assigned in the above rule 1 will appear in this new column. whereas the interaction of columns 3 (ab) and 6 (bc). pH. the pH!temperature interaction will appear in column 3 (ab). which were used to assign factors (BotSize. respectively). may be used to assign ﬁve factors (in columns 1.1a. will appear in column 5. 9. will be found in column 7. To decide which interactions should be computed for signiﬁcance. whereas scheme 4 is the design that may be used when all of the twofactor interactions of factor 1 with other factors are assumed to be important. this table can accommodate a maximum of 14 factors assigned to each of 14 columns. This fractional factorial design is based on the four-factor. respectively) in the design table. 9. The components are used to ﬁnd the columns where “column!column” interactions appear. literature data. VarLoc: fruit variety or locality. If a new factor is assigned to this column. Taguchi’s designs4 that are based on orthogonal arrays will be discussed due to ease of explanation of the principle and ANOVA (analysis of variance) computation. and any other sources should be taken into consideration. The other digits of 215 in the L16(215) designation show that there are 15 columns of two-level limit values (upper and lower limits represented by digits 1 and 2. 29Z512 experiments would have to be carried out. or their combinations. their grouping. with interactions designated as digits (interaction columns) on the lines. StTemp: storage temperature. Therefore. By assuming that some higher-order interactions are not important. Nine factors (WHard: water hardness. all of the available information derived from past experience. To use these components.7 many designs have been published. 8 and 15) along with interactions between each pair of factors.EXPERIMENTAL DESIGN AND RESPONSE-SURFACE METHODOLOGY 297 Since the early Yates’ algorithm using C and K signs that was employed by Box et al.1b. StTime: storage time.1 Two-Level Design Example A hypothetical example is shown in Table 9. its effect will be confounded with the interaction effect of the previously assigned two factors. Among the many of Taguchi’s designs for factors with two to ﬁve levels. 2. The bottom of the table shows the list of component rows expressed as alphabetical letters. PasTemp.. two-level full-factorial design. 6. PasTemp: pasteurization temperature.3 In this chapter. In theory. and Sugar: sugar content) are considered to be affecting the sediment formation in a fruit drink. the four-factor interaction of abcd. if this interaction should not be ignored. for example.3. BotSize: bottle size. StTime and BotShape. thus. LLC . An easier way of assigning factors is to use the interaction schemes shown underneath Table 9. Similarly.1a. it is q 2006 by Taylor & Francis Group. If all interactions are to be ignored. Several two-factor interactions were regarded to be nonsigniﬁcant: columns 5. the number of experiments required is 24Z16 (L16). If a full two-level factorial design is planned. 14. Four two-factor interactions (columns 3. the Plackett–Burman design.e. no new factor should be assigned to the same column. 4. when pH and temperature are assigned to columns 1 (component a) and 2 (component b). which is ab2c/ac. as shown in Table 9.2. Scheme 1. including one of the most popular. BotShape: bottle shape. which is abc. the interaction of columns 3 (ab) and 4 (c). the simple rules to be applied are: (1) multiply components of a combination of any two columns to which factors are assigned. (2) replace with unity when squares of letters appear in the component products.. a fractional factorial design with only 16 experiments may be used. and the interaction schemes coined as linear graphs are depicted underneath the table. i. To avoid overlooking important interactions. The schemes are composed of lines (interaction) connecting pairs of dots or “nodes” (for factor assignment).

q 2006 by Taylor & Francis Group. Tokyo. Design of Experiments. 9 > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > = 11 Source: From Taguchi.Table 9. LLC .1a L16(215) 298 Experiment No. 1. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Component 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 a 2 1 1 1 1 2 2 2 2 1 1 1 1 2 2 2 2 b 3 1 1 1 1 2 2 2 2 2 2 2 2 1 1 1 1 a b 4 1 1 2 2 1 1 2 2 1 1 2 2 1 1 2 2 c 5 1 1 2 2 1 1 2 2 2 2 1 1 2 2 1 1 a c 6 1 1 2 2 2 2 1 1 1 1 2 2 2 2 1 1 b c 7 1 1 2 2 2 2 1 1 2 2 1 1 1 1 2 2 a b c 8 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 d 9 1 2 1 2 1 2 1 2 2 1 2 1 2 1 2 1 a d 10 1 2 1 2 2 1 2 1 1 2 1 2 2 1 2 1 b d 11 1 2 1 2 2 1 2 1 2 1 2 1 1 2 1 2 a b d 12 1 2 2 1 1 2 2 1 1 2 2 1 1 2 2 1 c d 13 1 2 2 1 1 2 2 1 2 1 1 2 2 1 1 2 a c d 14 1 2 2 1 2 1 1 2 1 2 2 1 2 1 1 2 b c d 15 1 2 2 1 2 1 1 2 2 1 1 2 1 2 2 1 a b c d HANDBOOK OF FOOD AND BIOPROCESS MODELING TECHNIQUES Group (a) (b) (c) (d) > > > > > > > > > > > > > . 9 = (1) 2 3 13 6 5 4 (4) 2 1 14 (2) 15 14 13 3 2 7 8 (3) 7 3 4 2 (6) 8 10 11 13 15 9 1 3 2 7 5 6 10 1 13 9 8 4 11 15 14 5 4 9 15 7 12 9 1 11 5 6 10 6 12 10 12 8 (5) 1 3 1 15 14 13 5 7 9 11 12 4 5 10 8 4 5 7 6 3 12 15 14 13 12 2 8 10 9 11 14 . 1957. 9 > > > > > > > > > > > > > = > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > . Vol. G. Maruzen.

1c. can be assigned. e. At least one column should be left unassigned because it is required to serve for the error sum of square computation.1b Sediment Formation in the Fruit Drink Column WHard pH PasTemp StTemp StTime BotSize BotShape VarLoc Sugar Unused a a Interaction 1!2Zab/3 2!6Zc/4 7!9Zbcd/14 5!10Zabcd/15 1 (a) 2 (b) 6 (bc) 7 (abc) 9 (ad) 5 (ac) 10 (bd) 8 (d) 11(abd) 12 (cd).20 1. food colors and preservatives.05 0. Table 9. In the present example. and 14 (all are three-factor components) in Table 9.15 0.05 0. In total.00 0. 7. LLC .25 1. e.10 0.1c are carried out in random order to avoid carryover error. 11. and 8. The amounts of sediments measured as a result of each of the 16 experiments are also shown in Table 9. St temp PasTemp 3 Whard 13 12 Sugar StTime 15 14 pH BotShape Source: Nakai’s lecture note.EXPERIMENTAL DESIGN AND RESPONSE-SURFACE METHODOLOGY 299 Table 9.55 q 2006 by Taylor & Francis Group. VarLoc 4 BotSize recommended to assign new factors. 2.40 0..20 0. 13 (acd) Other factors.1a when other factors are already assigned to columns 1. possible interactions or an error term to columns for higher-level interactions.. 13.g.60 0. 16 experiments under the conditions shown in Table 9.50 0.50 0.g.85 0.1c The Amounts of Sediments Measured After Storage of Bottled Drink Exp 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 W Hard 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 pH 1 1 1 1 2 2 2 2 1 1 1 1 2 2 2 2 3 4 Bot Size 1 1 2 2 1 1 2 2 2 2 1 1 2 2 1 1 Pas Temp 1 1 2 2 2 2 1 1 1 1 2 2 2 2 1 1 St Temp 1 1 2 2 2 2 1 1 2 2 1 1 1 1 2 2 Var Loc 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 St Time 1 2 1 2 1 2 1 2 2 1 2 1 2 1 2 1 Sugar Bot Shape 1 2 1 2 2 1 2 1 1 2 1 2 2 1 2 1 1 2 1 2 2 1 2 1 2 1 2 1 1 2 1 2 12 13 14 15 Sediment -gr1.75 0. one experiment after another. columns 12 and 13 are left unused for this purpose.07 0.

i. ANOVA computation was performed in ordinary fashion to obtain sum of squares for treatment St. BotShape. the smaller the sediment.1) (9. and if all of the causes are correctly deﬁned. C 0:75 C 0:55Þ2 =16: All computations can be readily carried out using Excele functions SUM (total) and SUMSQ (sum of squares). It is interesting to note that the WHard!pH interaction is signiﬁcant although WHard itself is nonsigniﬁcant (Table 9. StTime..9 The ANOVA table thus calculated is shown in Table 9. At pH 6.1b). 9. despite the fact that highly trained skill. CT Z 1 (9. k X ST Z ðdataÞ2 KCT. and BotSize and WHard!pH interactions are less signiﬁcant (P!0. The speciﬁc combination of these two factors has an effect on the response. which in this case is the amount of sediment.3) where k is the number of experiments.1d:B).2.300 HANDBOOK OF FOOD AND BIOPROCESS MODELING TECHNIQUES Then. it is evident that pH is highly signiﬁcant (P!0. However. To extract useful information. Signiﬁcant interaction between pH and WHard is apparent in Figure 9. Any phenomenon is the consequence of its causes.1d. there should be no error introduced into the resultant data analysis. errors would always exist. such as those of Plackett and Burman. even after the pH!PasTemp interaction has been pooled into the error (panel A of Table 9. In the ANOVA thus repeated (panel B of Table 9. An example of St computation is: S t for bottle size in column 5 is ð1:25 C 1:00 C 0:10 C 0:05 C0:85 C 0:60 C 0:75 C 0:55Þ2C ð0:50 C 0:20 C 0:07 C 0:15 C 0:20 C 1:05 C 0:50 C 0:40Þ2 Kð1:25 C1:00 C .1a demonstrate a decrease in sediment as pH decreases and bottle size is increased. are pooled into the Error SS. The increase in sediment by increasing WHard is nonsigniﬁcant (Table 9. and BotSize!BotShape.” which would be true in the case of no interaction. it is humanly impossible to eliminate all errors throughout the entire experiment.2 Three-Level Design Example The two-level designs are useful for the purposes of screening inﬂuential factors. the sediment has increased even when the water is soft (Figure 9.05). The response curves illustrated in Figure 9.01).e. sum of squares for total ST. and use of instruments with high accuracy may diminish the size of error to the minimum. None of the nine factors is signiﬁcant in affecting the amount of sediment. in reality. The formation of sediment cannot be simply stated as “the lower the pH and WHard. but they are not adequate to represent response surfaces because they cannot illustrate concave or convex q 2006 by Taylor & Francis Group. Because of the importance of detecting factor–factor interactions. SS with F values lower than unity.1d). !2 1 k X data =k.3 many Latin square designs3 are also similar as far as undetectability of interactions is concerned. Therefore. This phenomenon is characterized by a great reduction of sediment at low pH of 3 when water is soft (W1). VarLoc. and correction term CT: St Z X 2 X 2 treatment 1 C treatment 2 KCT. these ineffective designs are not discussed further in this chapter.1b as the departure from parallelism of the response curves.3. There are many other fractional factorial methods aside from the Taguchi designs. appropriate knowledge.1d:B). LLC .2) (9. thereby increasing its degrees of freedom to 7 and resulting in an increase in the power of the analysis to detect signiﬁcant factors.1d).

13 0. “replace cubes (instead of squares) with 1.16 0.16 0.17 Df 1 1 1 1 1 1 1 1 7 15 MS 0.05 0.59 0.044 F 3.27 0.27 0. the relation of (ab2c2)!(ab2c)2/a3b6c4 becomes c (column 5).23 0.13 0.01)1.55 surfaces for which at least three-level designs are required.27 0.21 0. the relation of column 10 (ab2c2)! column 12 (ab2c)/a2b4c3 is used. with two exceptions.03 Column 1 2 6 7 9 5 10 8 11 3 14 15 4 4.09 0. this becomes a4b2. an interaction appears in two columns rather than one column.01)OFOF(0.EXPERIMENTAL DESIGN AND RESPONSE-SURFACE METHODOLOGY 301 Table 9.14 0. Source: Nakai’s lecture note. which turns to ab2 located at column 4.10b 3.94 6.21 0.3Z10.59 0. The example used here involves an experimental design to study the effects of four factors and three two-factor interactions on the solubilization of wheat ﬂour by heating with HCl (Table 9. Meanwhile. the component calculation to ﬁnd interaction columns is different from that used for two-level designs.00 0.01 0.01 0.16 0.25.13 0.04 0.68 0. b F(0.14 0.77 0.00 0.55 3.14 0.01].e.09 0.10b 2.27 0.2a is most suitable for this purpose. The solubility of the hydrolysates from each of the 27 experiments is measured as the responses..16 6. a FOF(0.16 0.04 0.04 0.05] 0. By the “replace cubes with 1” rule.86 1. The experimental conditions and the solubility obtained are shown in Table 9. [P!0.14 0. i.59.22 2. to ﬁnd the interaction of columns 10 and 12.00 1.09 0.21 0.31 2.05). For instance. Source: Nakai’s lecture note.32a 4. q 2006 by Taylor & Francis Group. this becomes a2b because b3c3/1. the interaction of column 10 and column 12 appears in columns 4 and 5.91 3.68 1.1.01).18 8.05 2. Taguchi’s L27(313) design shown in Table 9. F(0.04 0.59 0.05)1.13 0.073 F 2. [P!0.” Second. In summary.7Z12.05)1.61 13.74 2. LLC .21 0.17 Df 1 1 1 1 1 1 1 1 1 1 1 1 1 3 15 MS 0. when squared. 12. 13 SS 0. SS 0.05].27 0.09 0.27 0. First.1d ANOVA for Sediment Formation of a Fruit Drink Before (A) and After (B) Pooling Factors or Interactions Into the Error Term on the Basis of Mean Squares Factor/ Interaction (A) WHard pH PasTemp StTemp StTime BotSize BotShape VarLoc Sugar WHard!pH StTemp!StTime BotSize!BotShape [pH!PasTemp Error Total F(0.59 0.2c. The general strategies and procedures previously described for two-level designs also apply to three-level designs.2b). Factor/Interaction (B) WHard pH PasTemp StTemp BotSize Sugar WHard!pH StTemp!StTime Error Total F(0.27 0.27 0.14 0.7Z5.

20 0.80 0.50 0.30 0.20 2.90 W1 W2 0.00 Water hardness 2.00 0. ml (a) (b) Hard 0.70 1.90 0.00 Figure 9.00 0.00 pH 1.30 0.30 0.00 0.60 0.40 0.90 0.00 6.1a Response curves of sediment formation in the fruit drink as a function of (a) pH.70 0.50 0.00 0. gr 0.50 0.00 4.90 0.70 Sediment.00 3. gr 0.50 Soft 0.60 0.10 400 500 600 700 800 900 1000 1100 Bottle size.1b Interaction effects on sediment formation.60 0. LLC 1.10 0 2 3 4 pH Figure 9.00 0.80 0.40 0.00 7. (b) bottle size.40 0. 2 = hard) pH 6 pH 3 .302 HANDBOOK OF FOOD AND BIOPROCESS MODELING TECHNIQUES 1.10 0 5 6 7 0 1 2 3 Water hardness (1 = soft. gr 0. 1.00 0.20 0.00 (c) 1.60 0.40 0. and (c) water hardness.30 0.80 Sediuent.20 0.90 0.80 0.80 Sediuent.60 0. q 2006 by Taylor & Francis Group.70 5.30 0.50 0.70 0.40 0.

2b.3) described for the two-level design example can again be used. and Se is SS for error. 9 > > > = Group (a) (c) (1) (2) 3. Design of Experiments. The conﬁdence limit (CL) is calculated using the following formula: h pﬃﬃﬃipﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ CL Z tð0:05Þ at dfe = k Se =dfe (9.EXPERIMENTAL DESIGN AND RESPONSE-SURFACE METHODOLOGY 303 For ANOVA computation.2d. Tokyo. k is the number of data. Vol. Table 9.13 2 5 8 11 2 8. the same equations (Equation 9.4 6. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 Component 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 3 3 3 3 3 3 3 3 3 a 2 1 1 1 2 2 2 3 3 3 1 1 1 2 2 2 3 3 3 1 1 1 2 2 2 3 3 3 b 3 1 1 1 2 2 2 3 3 3 2 2 2 3 3 3 1 1 1 3 3 3 I 1 1 2 2 2 a b (b) 4 1 1 1 2 2 2 3 3 3 3 3 3 1 1 1 2 2 2 2 2 2 3 3 3 1 1 1 a b2 5 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 c 6 1 2 3 1 2 3 1 2 3 2 3 1 2 3 1 2 3 1 3 1 2 3 1 2 3 1 2 a 7 1 2 3 1 2 3 1 2 3 3 1 2 3 1 2 3 1 2 2 3 1 2 3 1 2 3 1 a c 8 1 2 3 2 3 1 3 1 2 1 2 3 2 3 1 3 1 2 1 2 3 2 3 1 3 1 2 b c2 9 1 2 3 2 3 1 3 1 2 2 3 1 3 1 2 1 2 3 3 1 2 1 2 3 2 3 1 a b c 10 1 2 3 2 3 1 3 1 2 3 1 2 1 2 3 2 3 1 2 3 1 3 1 2 1 2 3 a b2 c2 11 1 2 3 3 1 2 2 3 1 1 2 3 3 1 2 2 3 1 1 2 3 3 1 2 2 3 1 b c2 12 1 2 3 3 1 2 2 3 1 2 3 1 1 2 3 3 1 2 3 1 2 2 3 1 1 2 3 a b2 c 13 1 2 3 3 1 2 2 3 1 3 1 2 2 3 1 1 2 3 2 3 1 1 2 3 3 1 2 a b c2 > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > . q 2006 by Taylor & Francis Group. 1. the only difference is that the treatment levels are 1–3 instead of 1–2.2a L27(313) Experiment No.4) where t (0. G.05. 9 > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > = > > > .2a and Figure 9. Maruzen.1 0 12 .7 9. The ANOVA table is shown in Table 9.4 6. LLC . The response curves for the main effects and interactions are shown in Figure 9. respectively.11 5 Source: From Taguchi. 1957.05) is the Student’s t value at PZ0.7 9 10 12 13 1 3.1 through Equation 9.

When autoclaved (120 8C for 15 min). but yielding solubility responses that are not parallel to those at other HCl!temperature-time combinations.4 HCl 8.11 6.8 87.5 85.9 85.9 88.1 91.304 HANDBOOK OF FOOD AND BIOPROCESS MODELING TECHNIQUES Table 9.7 84.0 68.6 80.” However.7 H R 10 12 13 The parallel curves for ﬂour-HCl (Figure 9. 12 (ab2c).3 76.4 94.4 92.5 89. LLC .4 71.2c Solubility of Acid-Treated Wheat Flour Experiment 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 F 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 3 3 3 3 3 3 3 3 3 HCl 1 1 1 2 2 2 3 3 3 1 1 1 2 2 2 3 3 3 1 1 1 2 2 2 3 3 3 H 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 R 1 2 3 2 3 1 3 1 2 2 3 1 3 1 2 1 2 3 3 1 2 1 2 3 2 3 1 Solubility (%) 95.2 85.2b) indicate no signiﬁcant interaction between these two factors.1 94. almost the same high solubility was obtained at all HCl concentrations studied.9 94.0 94. Table 9.2 95. 13 (abc2) Interaction 1!2Zab/3C4 2!5Zbc/8C11 1!5Zac/6C7 F 3. and the effects of these factors can simply be stated as “the higher the HCl concentration.3 q 2006 by Taylor & Francis Group. the HCl ! temperature-time interaction is signiﬁcant.7 80. the higher the obtained solubility at all ﬂour concentrations.0 86.4 81.9 94.0 85.8 92.4 95.0 83.2b Wheat Flour Solubilization Factor Flour (F) HCl Heat treatment (H) Reductant (R) Unused Column 1 (a) 2 (b) 5 (c) 9 (abc) 10 (ab2c2).

53.2a Response curves of acid-solubilized wheat ﬂour as a function of percent ﬂour.04 88. % 97.05 0.00 12.60 38. F(0. P!0.50 80.43. q 2006 by Taylor & Francis Group.00 2.46 497.91 995. F(0.40 9. HCl concentration and temperature-time treatment.2d ANOVA of Acid Hydrolysates of Wheat Flour Factor/interact ion Wheat ﬂour HC1 Temperature–Time Reductants Flour!HCl Flour!T–T HCl!T–T Error Total SS 135.05)2.50 85.51 22.50 80. N 0. a P!0. % 90.14.50 75.00 10.20 Flour.10 2.00 0.53 1442.69a 121.20 6.50 90.50 85.77 0.66a 0. b 92.50 Solubility.00 77.96 Df 2 2 2 2 4 4 4 6 26 MS 67. LLC .50 95.6Z4.00 Solubility. % 87.25 1.15 0.00 82.58a 18.60 3.05)4.00 0. 13 0.00 100° 40 min 100° 120 min 120° 15 min Temperature-time treatment Figure 9.00 6.43 b F(0. Source: Nakai’s lecture note.20 4.EXPERIMENTAL DESIGN AND RESPONSE-SURFACE METHODOLOGY 305 Table 9.00 8.6Z9.6Z5.09 F 16.33 5.79 24.50 92.01)2.15.83 76.10 HCI.6Z10.05.00 4. F(0.64 152.01.00 82.00 87.01)4.

can be ignored as in the case of the Taguchi method10 is debatable. which are difﬁcult to change the experimental conditions. 11. and 14) can be used for assigning factors. i.15 min Flour 5% 7% 11% 80. eight columns (1. they do not identify unspeciﬁed interaction effects that are confounded with the main effects and the speciﬁed interaction effects. Our result using the Taguchi method identiﬁed signiﬁcance of the 1!2 interaction appearing on column 3 (Table 9.00 70. The three-level designs such as L27(313) are valuable for approximating the response surfaces.16 HCI. The “interaction graphs” of Kacker11 may circumvent these problems. 8. Also. columns 1 and 2 belong to different groups.16 HCI. they do not provide information on confounding relationships. column 3 consisted of four interaction effects of 1!2. Assigning these two factors to columns in the same group is not recommended due to less chance of useful interaction effects compared to the main effects of the already assigned factors.10 0. and 13!14.00 120°C . whereas the rest (3.3. 6.10 0. 9 (ad) and 10 (bd) may be a problem because the 2 (b)!7 (abc).02 0. which is similar to the algorithm of backward stepwise multiple regression. Second. may confound with those four factors. 1 (a)!8 (d). especially in biotechnology. However.06 0. 4!7. especially higher than second order.00 0. as well as 1 (a)! 11 (abd) and 2 (b)!8 (d) interactions.02 0.” respectively. especially for factor screening. N 60. according to the interaction graph of Kacker. therefore. 13. It was reported that the linear graphs of Taguchi had two disadvantages. may improve the screening efﬁciency of factors.14 0. LLC .. x 80. More efﬁcient designs than the L27(313) design for the purpose of response-surface modeling will be discussed in the following section.00 70.06 0.04 0.1a).08 0. Because factors were not assigned to columns 4 and 13.04 0.2. group “a” and group “b.2b Response curves of acid-solubilized wheat showing signiﬁcant (right side) and nonsigniﬁcant (left side) interaction between factors. According to the interaction graph for L16(215).00 100°C . the pooling-up process of the resultant ANOVA. (Table 9. 8!11. 6 (bc).e.3 How to Select Designs The orthogonal array of Taguchi4 was constructed on the basis of a probabilistic rule that two factors. 4. 12. 9. whereas the 8!11 q 2006 by Taylor & Francis Group. the above-mentioned 4!7 and 13!14 do not eventually exist in this example.00 Solubility.14 0. 10. 2.1b. We have found that Taguchi’s L16(215) design is the most efﬁcient and useful for general purposes. Whether higher-order interactions. and column 3 was left blank to calculate the 1!2 interaction effect.11 First. 1 (a)!7 (abc).1b). and 2 (b)!11 (abd) interactions.12 0.00 0. high-degree interactions may not be always ignored.306 HANDBOOK OF FOOD AND BIOPROCESS MODELING TECHNIQUES 100.00 100. and 15) are for assigning two-factor interactions or to be left blank for error SS computation.12 0. respectively.08 0.40 min 90. 7. 5. N Figure 9. In practice. In the multivariate era. should be assigned to two columns belonging to different groups in L 16(215) design. In the case of the example of fruit drink sediment in Table 9.00 100°C -120 min 90. assigning factors to columns 5 (ac). 9.1d:B and Figure 9.

New York. 7.3. Theoretically. the factors computed to be signiﬁcant even after confounding with interaction effects should be truly signiﬁcant. 2(b). 2. even factors assigned to one-letter columns 1(a). abd. The important fact is that the controlling effects of main factors are always vulnerable to the inﬂuence of any two-factor interaction. This is especially true in the case of factor screening as a prerequisite of the subsequent optimization. 2. 4. This is important for factor screening. 13. the Taguchi’s linear graphs were made for detecting the most probable interaction effects. 5. thereby making it difﬁcult to separate individual interactions. abc. 6. b. McGraw-Hill. respectively. and 14 all bear one or three letters as components: a.. which belong to different groups. whereas column numbers outside of the parentheses must be used for assigning factors ﬁrst. On the other hand. 1996. Columns 3. Whether the two-level interactions adversely affect the main effects or not should be tested as was done in the ANOVA computation shown above. He recommended a simpler digital series of 1. bd. signiﬁcance in column 3 representing interactions should be mostly due to contributions of the interaction of factors assigned to columns 1 and 2. assigning factors to these columns is reasonable. 4(c). Column numbers in parentheses may be assigned in any order. as illustrated in the fruit drink sediment example. 10. According to the two-level interaction table shown in Table 9. 4. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 2 3 — — — — — — — — — — — — — 3 2 1 — — — — — — — — — — — — 4 5 6 7 — — — — — — — — — — — 5 4 7 6 1 — — — — — — — — — — 6 7 4 5 2 3 — — — — — — — — — 7 6 5 4 3 2 1 — — — — — — — — 8 9 10 11 12 13 14 15 — — — — — — — 9 8 11 10 13 12 15 14 1 — — — — — — 10 11 8 9 14 15 12 13 2 3 — — — — — 11 10 9 8 15 14 13 12 3 2 1 — — — — 12 13 14 15 8 9 10 11 4 5 6 7 — — — 13 12 15 14 9 8 11 10 5 4 7 6 1 — — 14 15 12 13 10 11 8 9 6 7 4 5 2 3 — 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 Source: Adapted from Ross. c. a fractional factorial design cannot avoid the confounding of multiple interaction effects. 7. Taguchi Techniques for Quality Engineering. Therefore. (3. should be left blank for computing two-factor interactions. and the impact of the two disadvantages mentioned above may therefore often be negligible with no practical relevance. Column no. respectively. seven interactions may be confounded. 2nd ed. (11. acd.11 interactions consist of at least four interactions. 11. 8. q 2006 by Taylor & Francis Group. 14) is recommended. it is worth noting that the great saving in the number of experiments that can be gained by sacriﬁcing some information on the interactions using Taguchi’s orthogonal array method is still extremely valuable. bc. and 12. 8. 9. At least column 15 should be left blank for error computation. d. P. for column 3. 13. Recently. 4. and cd. However. 5. Ross10 proposed new interaction tables by thoroughly covering possible interactions. According to the interaction graphs of Kacker. Columns 1.. may each be confounded with seven two-level interactions each.3 Two-Level Interaction Table Column no. 9. 12) for assigning more than eight factors. Table 9. and bcd. 10. 11. 7. 13. and 8(d).EXPERIMENTAL DESIGN AND RESPONSE-SURFACE METHODOLOGY 307 interaction can also be considered nonexistent because of the near-zero sum of squares for VarLoc assigned to column 8. 6. it is recommended to follow this order of selecting columns. No matter which design is used. a shorter digital series of 1. 2. ac. if one has 6–8 factors to be assigned. corresponding to ab. J. As the number of factors to be assigned increases. For instance. 14. ad. Other unassigned columns are left blank to use in computation of interactions or error. 8. If we accept low probability of signiﬁcance of three-factor interactions. LLC .

The earliest technique of RSM requires central composite designs to draw 2D or 3D response surfaces.4. a random search is the only powerful approach in terms of overall efﬁciency to target the global optimum.0. However. LLC . Under this circumstance. A 2D map is used to visualize the response surface for each factor because it is not feasible to visualize multivariate response surfaces with greater than three dimensions. Since our presentation of RCO. site-directed mutagenesis. modeled using quadratic and/or polynomial factorial equations. to search for the global optimum.1 Introduction RESPONSE-SURFACE METHODOLOGY The basic procedure of the experimental designs is to compute ANOVA using factorial analysis. Recent examples of applications of factorial and fractional factorial experimental designs in food/bio processes are shown in Table 9. and (2) “evolutionary operation” and its derivatives to rotate the search toward the optimum during iterative search. There have been two schools of thought with regard to approaches for approximating the response surface and ﬁnding the optimum. the subsequent aim becomes to ﬁnd the optimum that appears on the response surfaces. nonlinear global response surfaces.4. SSO (sequential simplex optimization)... and these methods are usually simpler in experimentation than those in school 1. it was recommended to repeat the SSO process by initiating the search from different starting locations within the global search spaces. which belongs to school 2. etc. Furthermore.” Subsequently. At the same time. namely (1) extension and/or continuation of experimental designs followed by curve-ﬁtting to illustrate the response surfaces. a stepwise approach to ﬁnd the optimum was proposed.4 Recent Applications of Experimental Designs in Food/Bio Processing Method Full factorial Fractional factorial Plackett–Burman Taguchi–Orthogonal a Food/Bio Processes MAPa of sponge cake Finding Minimum K sorbate to prevent fungal spoilage Routine control of chocolate bars Macerating enzymes from transgenic wine yeast Reference 12 Triglyceride GC of cocoa butter equivalents Red wine aroma 13 14 Modiﬁed atmosphere packaging. despite its speculative property and being basically inefﬁcient in nature.308 HANDBOOK OF FOOD AND BIOPROCESS MODELING TECHNIQUES Table 9. such as for the purposes of mixture designs.4 9. We have proposed “RCO (random centroid optimization)”—a randomized version of SSO with an additional. food formulation. characteristic 2D mapping—to approximate the true response surfaces. the visualization of the curvature of response surfaces requires more response values per factor than provided by three-level experiments. After ﬁnding signiﬁcant factors by these experimental designs. i. In the life sciences. container designs.e. 9. the curve-ﬁtting technology currently available is limited in ability to depict rugged. The incipient design of methods in school 2 can be rather simple. a variety of RCO versions have been developed. mixture designs were developed to meet the constraint requiring that “the sum of components constituting a food formula is 1. q 2006 by Taylor & Francis Group. Accordingly. fractional factorial designs are essential for efﬁcient analysis in terms of costs and labor for biological experiments.

the number of experiments. a 2k factorial design (with design points at G1). These extended applications will be discussed later in this chapter. then nZ81.92 14. 6. where “1” is the center point that is usually conducted in replicate to provide an estimate of error.76. 3.. K1. i. The minimum number of design points in a central composite design is 2kC2kC1. central composite designs are recommended for RSM ﬁtted to second-order polynomial models.20 25.3). for example.682 0 0 0 Time K1 K1 K1 K1 1 1 1 1 0 0 0 0 K1.682 1. 27. The general form of a-values is: a Z 2k=4 .22. heating temperature. with the recommended 5.65 0. with a-value of 1. Central composite designs are a “composite” of a star design (with axial points at Ga). 19.682 1. 9. respectively. For example.99 24. The combination of the two designs leads to ﬁve levels of each factor. 0. 24. becomes unrealistically large.e. n. although a 3k full factorial design could be satisfactory to generate the necessary data to ﬁt a second-order polynomial model. For example. and 4.5 and Figure 9. LLC . 20. pH.682 1. with [Ka. or higher-order polynomial.5) where k is the number of variables.90 4.53 to estimate error variability. (9. the usual number of experiments for kZ2. for kZ4. 22.10 6.EXPERIMENTAL DESIGN AND RESPONSE-SURFACE METHODOLOGY 309 with potential applications in many other areas of food and biological processes. q 2006 by Taylor & Francis Group. For a model to account for a qth order effect in a factor.682 (Table 9.37 11. and Ca] coordinates expressed as coded values. Thus.49.83 19.88. and a common center point of the two designs (at zero). for many life science experiments. and 7 replicates of the center point. For this reason. second-order.68 17. the response is not ﬁrst order.03a Average of hexaplicate: 22. However. C1.682 0 Yield (g) 16. are 13. usually showing some curvature and factor interactions or second-order relationships. and 27.27.44 12. For such cases.4. the experimental design must have qC1 different levels in that factor.682 0 0 0 0 0 Temperature K1 K1 1 1 K1 K1 1 1 0 0 K1. The example used here is the optimization of product yield by changing three factors.5 Product Yield Experiments Using Three-Factor Central Composite Design Experiment 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 a PH K1 1 K1 1 K1 1 K1 1 K1. and time. Table 9.16 7.90 7. and 31 experiments.2 Central Composite Designs The minimum number of experiments necessary to estimate the parameters in a responsesurface model depends on both the number of factors (k) and whether the model is expected to be ﬁrst-order.39 18. the experimental design for a ﬁrst-order model with k factors could be either a 2k full factorial design or a two-level fractional factorial design.50 16.

all possible q 2006 by Taylor & Francis Group. i.19 g. M.0) only.099 for x1.6) with R2Z0.e. K0. M. Weinheim. 608 and 100 8C (Temp) and 1 and 10 min (Time).. 9. extreme vertices designs using limit or bound values (0 and 1. In the optimization of mixtures or food formulations. pp 759–773. respectively..) The limit level values corresponding to the coded values of K1. 1998. Otto. x2. the optimal scale or coded values were computed to be K0. points are spread evenly over the whole simplex factor space and lattice. (9.682 were assigned as follows: 4 and 9 (pH).682. Then. Mermet. J. or 100%.355. These values are equivalent to the factor conditions of pH 5. Ed.392. an ordered arrangement of points may have a special correspondence to a speciﬁc polynomial equation. In this chapter. J. By solving the simultaneous equations resulting from derivatization then equalizing to zero. without selecting intermediate values (fractions such as 0. (Adapted from Otto. The most well-known designs for ingredient mixing are simplex-lattice designs. and Widmer. Analytical Chemistry. will be discussed..3 Mixture Designs Mixture designs are a technique to ﬁnd the combination of the food ingredients that yields the best quality of a food product. M. and x3. with the estimated best yield of 25. Wiley-VCH. Kellner. the sum of all the ingredients must always be one.15 To accommodate a polynomial equation.8 8C for 5. The limit values (bounds) of each ingredient are initially selected.3).76 min. LLC .310 HANDBOOK OF FOOD AND BIOPROCESS MODELING TECHNIQUES +1 x3 0 +1 −1 −1 0 x1 +1 0 −1 x2 Figure 9.923.4.... The prediction equation obtained by curve-ﬁtting using multiple linear regression analysis was: 2 2 y Z 24:04 K4:72x1 K1:189x2 C 0:229x3 K5:426x2 1 K0:884x2 K5:159x3 K1:674x1 x2 K1:146x1 x3 K0:971x2 x3 . R.. Mixture designs must therefore consider this constraint that changing the proportion of one component automatically leads to a change in the proportion of the other components. -M.682 and C1. the simplest mixture design method.3 Central composite design for three factors. *aZ1.53 and heating at 75. and 0.

85 0.0 (100%) is assigned to the remaining qth ingredient.02 0.20 0.3 9.02 0.7) For instance.8) with R2Z0.03 0.025 Measured 7 6 15 14 13 10 18 20 10 12 16 17 13 Predicted 6. as the total was 1.20 0.10 0.70 0.0 13. x1.EXPERIMENTAL DESIGN AND RESPONSE-SURFACE METHODOLOGY 311 combinations of the bounds for q ingredients are computed by selecting bound values of qK1 ingredients at a time.03 0.981. this combination is withdrawn from the design.77 0.6. Surimi base.20 0.7 10.2 6. potato starch.02O1.10 0.78 0. a combination of vertex values of 0.6 Extreme Vertices of Formulation of Kamaboko Gel strength (N) Base 1 2 3 4 5 6 7 8 9 10 11 12 Center 0.13 0.03 0.8 by eliminating the intercept and square terms from the full factorial equation exempliﬁed in Equation 9. starch.10 0.4 12.07 0.70 0.02 0.70 0.6 8.02 0.03 0.17 0.7 14. if not. the maximum number of design points for a q-component design will be q$2qK1.08 0.10 0. Table 9.0 13.1.10 0. where y is peak force (N).0 leads to Equation 9.85 0.02 for the four ingredients was eliminated from Table 9.20. and their bounds were set as follows: 0:70 % Base % 0:85 0:10 % Starch % 0:20 0 % Water % 0:10 0:02 % Salt % 0:03 (9. water.6.03 0.7 15. In practice.0. The value subtracted from 1.02 0.10 0.03 0. Note that the substitution using the composition constraint x1Cx2Cx3Cx4Z1. (9.2 12. and salt were the ingredients used. 0.70. and 0.5 13.12 0. water. An example is shown in Table 9. x3. respectively.15 Water 0. LLC . many points will be eliminated since the remaining qth component may exceed its bounds.03 0.02 0.6 for maximizing gel strength of ﬁsh cake (kamaboko) measured as the peak force of Instron trace (N).85 0.77 0. x2.20 0.18 0.3 q 2006 by Taylor & Francis Group. 0.10 0 0 0 0.85 0.775 Starch 0. and salt. In theory.70 0.05 Salt 0.10 0 0.02 0. provided it is within its bounds.78 0. A quadratic model ﬁtted to the data was: y Z 9:783x1 K508:93x2 K378:34x3 K3987:0x4 C 629:55x1 x2 C 265:92x1 x3 C 3827:28x1 x4 C 1318:19x2 x3 C 7177:21x2 x4 C 4977:35x3 x4 . and x4 are base.7 15.

8% potato starch. When the search starts from point A on the x-axis by moving parallel to the y-axis. as illustrated by R in Figure 9.16 Thus.4 Sequential Simplex Optimization (9. 19. thus accurate prediction cannot be expected. Spendley et al. the section proﬁle Sb also shows that point C on the xy plane is the highest point (so-called ridge problem). P. if the search instead reaches a local hilltop D (a local optimum). This ﬁgure shows that the iterative or evolutionary move of the triangular simplex ﬁnally has reached the summit.. there is no chance to move away from it as long as only one-search-at-a time processes parallel to x and y axes are alternately used. and 4. 265:92x1 C 1318:19x2 C 0x3 C 4977:35x4 C l Z 378:34. multiple linear regression analysis such as used here could not prevent the boundary violation for salt constraint of 2–3%.9% salt.9% is obviously a violation of the search-space boundary. the probability of reaching the true optimum is diminished when the response surface contains a ridge (interaction). point C appears to be the location of highest point on section proﬁle Sa. that basically searches for the optimum (either a maximum or a minimum point) by sequentially moving away from the worst response or point in a small search area. 1% water.5 N using 74. LLC .312 HANDBOOK OF FOOD AND BIOPROCESS MODELING TECHNIQUES The simultaneous equations to be solved were: x1 C x2 C x3 C x4 Z 1:0. The well-known critical rule of polynomial surface derived from curve ﬁtting is that the accuracy of approximation is restricted to values within the range of the search spaces. including the top equation for composition constraint.6. This predicted value is in fact lower than the actual value of 20 N measured for experiment 8 in Table 9.10) j / / / / /j jq q pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃÉ À pﬃﬃﬃÁÈ É À pﬃﬃﬃÁÈpﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ where p Z 1= k 2 ðk K1Þ C k C 1 k C 1 K1 . termed a simplex. 629:55x1 C 0x2 C 1318:19x3 C 7177:21x4 C l Z 508:93.2% surimi base.4.4. Furthermore. 0x1 C 629:55x2 C 265:92x3 C 38:28x4 C l ZK 9:783. 5) to that of the above simultaneous equations. 9. The p and q values for and q Z 1= k 2 q 2006 by Taylor & Francis Group.9) Sequential simplex optimization is an empirical feedback strategy of optimization.17 recommended the use of the following (kC1)!k matrix for initiation of SSO with k factors: j0 jp jq 0 q p 0 / 0j q / qj q / qj q / pj (9. The salt content of 4. However. Simplex evolutionary operation owes most of its success to its ability to handle multiple variables including mutual interactions. The greatest gel strength that can be expected was calculated to be 15.e. These problems could be a drawback of curve-ﬁtting optimization in comparison to an evolutionary operation approach. 3827:28x1 C 7177:21x2 C 4977:35x3 C 0x4 C l Z 3987:0: where l is a Lagrangian multiplier to match the number of independent variables (i. Perpendicular to the x-axis.

6. respectively.0 E: Expension. The response value of R vertex (Yr) is compared as shown in the ﬂow chart (left half of Figure 9.4 Hypothetical illustration of ridge problem.5 Flow chart of the simplex minimization (left side). 0.943. when the lower and upper limits of each factor are designated as 0 and 1. k = 0. k = -0. This cycle is iterated by moving to oblique directions (not vertical direction) until reaching a summit. k = -0.25 CW' : Massive contraction of W.25 Y stands for the response C ðP K Figure 9.5 (right half). the second simplex is formulated. respectively.912.0 Cr : Contraction of R. By replacing W with the best vertex selected from the ﬂow chart. k = 2. k = 1.5 CW : Contraction of W.259). After carrying out the experiments as designed above. (0. Calculation of the new vertex using V Z P W Þ (rightside). LLC . 0. An example of computing the second vertex in simplex search is shown in Figure 9. local optima. k = 0.966.236).5 Cr' : Massive contraction of R. the reﬂection vertex R is computed using C ðP KW Þ where W is the worst vertex and P is the centroid vertex that is the average of all RZ P vertices within the simplex except W as shown in Figure 9.EXPERIMENTAL DESIGN AND RESPONSE-SURFACE METHODOLOGY 313 Z Y P D B Sa Sb C O R A X Figure 9.926. 0. and 5 are (0. 3. q 2006 by Taylor & Francis Group. 0.219) and (0. kZ2. R YR<YB NO YR<YN NO YR<YW NO CW YES E YE<YR NO YES E E N Cr R YES R Cr W P B YES Cr YC >YR r YES Cr' YC >YR r' YES Cr NO Cr' Replace W NO YC >YW r YES CW' CW' CW NO W: Worst vertex B: Best vertex N: Next-to-the-worst vertex P: Centroid of N-B vertices R: Reflection.5). 4.205). and simplex search for the global optimum. (0.

73+(74. * shuryo.4 72.4 4.9 84. a new approach of RCO was proposed.4 2.18 Each search cycle in the RCO consists of a regulated random search.2 As response YE<YR R replaces W Simplex 2 constructed: Vertex 5 2 3 4 B 9.73−70) As response YR>YB.4 Time 11.5 6.4 79. where n can frequently be greater than three for four factors. Initial simplex and recovery obtained: Inferiority order W1 W3 W4 W2 Vertex 1 2 3 4 B (%) 0 9. which will ruin the optimization efﬁciency.6 Model computation of vertices in simplex search. whereas the SSO requires 5!n experiments.ca.4 2.4 11. expansion is tried Figure 9. This computer program written in Quick Basic is available on request.4+2. The ranges set for search are 0–10% and heating at 70-80˚C for 2–12 min. q 2006 by Taylor & Francis Group. without any warranty of homing in on the global optimum. for optimization involving four factors.4 Recovery (%) 60 W 72 75 B 68 Centrold P= (9.4.5 4.5 Expansion E = = 18. LLC .4+2. compared to (kC 1)!n in the case of SSO.* 9.4 Temp 79. The number of experiments required is 2kR9 for the random search and generally less than four for the centroid search. with a random search replacing the simplex search.nakai@ubc.4)/3 = 4.4 2. and requires much less number of experiments than that of the SSO.5 Random-Centroid Optimization Because SSO cannot guarantee ﬁnding the global optimum unless the simplex search cycle is repeated several times after shifting the search space. a centroid search and mapping.73 11.73−2) 16.4 74.4 11. Therefore.314 HANDBOOK OF FOOD AND BIOPROCESS MODELING TECHNIQUES Problem: The recovery of product A is optimized by varying the amount of additive B and heating temperature and time.4 Temp (˚C) 70 72. with a slightly reduced number of experiments for every succeeding cycle.73 = 9.4 Time (min) 2 4.5 9.2 76 (Vertex 6) 80 (Vertex 5) Reflection R = 74.5 72.46 6.73+2x(6.4 2.73 79. a maximum of 13 experiments is required for the ﬁrst cycle of RCO.4 4.4 79. where n is the number of replications of the SSO cycle until an optimum is reached.4 72.

8 0. This requirement is in a good agreement with the above-mentioned symmetry rule for factorial designs. whereupon Absorbance × 280 nm 1. x2: NaCl.8 0.2 Absorbance × 280 nm 0. (4) the response values of the experiments are recorded. New search spaces different from the current spaces and with usually narrower search spaces are determined based on these maps and used as the new search spaces in the subsequent cycle. The routine procedure for cycle 1 of RCO is as follows: (1) The factors and their search limit values are entered into the RCO software. LLC .. x5: time. The maps drawn after every cycle are essential in deciding the direction of move for searching toward the global optimum. x4: temperature. and Funane. Ogawa. Nakamura.4 0 15 20 25 30 35 40 X3X4 1. and (3) the experiments are carried out.2 0.4 0 10 15 20 25 30 35 40 Time (min) 45 50 55 Figure 9.. An example of the effects of factor-ignoring on the mapping is shown in Figure 9. NaCl and time were not ignored (left half) and ignored (right half). Clearer trends towards the potential location of the global optimum are observed after using the factor-ignoring processes.7 Effects of factor ignoring on maps of enzymatic activity (A280). S. S.) q 2006 by Taylor & Francis Group.8 0.. a single. To extract the maximum available information for deciding on the search spaces of the subsequent cycles.. M. Dou. J..or double-factor ignoring process was introduced. 2003. 25–47. in the middle row for the response surface of temperature.EXPERIMENTAL DESIGN AND RESPONSE-SURFACE METHODOLOGY 315 The random design in each cycle of the RCO is regulated so that at least one experiment is located in each quartile within the search space to assure near-even distribution of the search. International Journal of Food Property.7.4 0 100 150 200 Pressure range (Vpa) 250 X2X5 1. x3: pressure. x1: reactant concentration. (From Nakai. (2) the random design of experiments is printed out.18 which illustrates the results obtained with (right half) and without (left half) the factor-ignoring process. K. For instance.2 Absorbance × 280 nm Temperature (°C) 0. 6.

2 Food Formulation Factorial analysis. speciﬁcally bottom width.22 9. 9. The package includes instructions on how to use the programs and also optimization model equations for practice training.23 which was then used as an example of market survey. 3. the RCO was demonstrated to be successful in ﬁnding the global optimum.4. mixture designs. compared to less than 50 iterations without being stalled when the RCO was used. when an unexpected amyloidosis problem resulting in insolubilization during isolation from yeast cells reduced the papain inhibitory activity of the isolated cystatin. An example is shown later in the application of RCO to genetic modiﬁcation studies of human cystatin C. The optimal preference of glass shape for various consumer groups was determined using this RCO shape optimization.e.agsci.21 Also. an example was shown in Nakai et al.ubc. but rarely used in general optimization purposes in food and bioprocess applications. top diameter of the hand-holding portion. the original RSM. RCG (RCO for genetics) found an explanation of the potential mechanism of this phenomenon. the Taguchi method.20 The component constraint of SCiZ1. The simultaneous shift included in the RCO package was useful in the model optimization trials for ﬁne tuning of the search. and diameter of the round top of the glass shape. the response values are reported and the summary data are printed. Using this analysis.0. for the ﬁrst time. This cycle is continued using the new search spaces in the subsequent cycles. i.4. where i is the number of components C. cycles 2. in which experimental designs become more complicated along with enhanced unreliability as the number of components increases.3 Shape Design and Market Survey Glass shape was optimized by changing elemental measures.6. and constrained simplex optimization were compared for food formulation purposes. In comparison to the conventional mixture-design technology.4. RCO and RCG can be downloaded from ftp://ftp. LLC .4. Even if the SSO search did not stall. height. q 2006 by Taylor & Francis Group. the best ﬂuorescent-probe method for protein surface hydrophobicity was. was imposed..6 Extended Application of RCO Extension of the applications of RCO for diverse research problems in food science and related disciplines are illustrated in the following examples. (6) the computer program immediately moves to the mapping process from which new search spaces are selected. 60–180 iterations were needed to reach the global optimum.6. Furthermore. formulation of a food with component constraints was successfully performed using RCO by incorporating a penalty function accommodating component constraints.ca/foodsci/. (5) after conducting the centroid experiments. RCO could readily overcome this problem due to its evolutionary operation nature.316 HANDBOOK OF FOOD AND BIOPROCESS MODELING TECHNIQUES the centroid design is printed out. 9. the search was stalled at local optima ﬁve times during 20 optimization runs. the RCO is advantageous with a possibility of making a new discovery due to its speculative nature.18 When the SSO was applied to a Fletcher-Powell arctangent model containing two local minima in addition to the deeper global optimum in between. Based on countless trials for many model computations.6.1 Mixture Designs The RCO approach was applied for a mixture design: the analysis of components of Raman spectra to best ﬁt to the surface hydrophobicity of proteins determined by different ﬂuorescent probes. selected based on quantitative assessment. etc. 9.

1 ND 1.07e 2.97G0. there is no limit for the number of factors manipulated by the RCO program.16e ND 2.75G0. and to simultaneously Table 9.08 0.1 EI9R/H86F 68.15G0.4 G32Y/A95D 69. Kitts.30G0.56G0. Nakamura...0 68.7 V10S/R93G 67.43e 2. H. 2002.13 0.08 0. E19R L27S S2K D15P G12W A16W L9V E20K R8N D15S PI3F L9F E2IS R25Y R8A V10S NDd 72..3 66.11 2. such as shape or color combinations.12 1. centroid search R8A/S98G 63.19G0. a T ⁄ (8C).00G0.6.5 G12W/H86V 70. although the RCO program currently available has a set limit to restrain optimization computation within the memory capacity of commercially available PC computers.1 60.. half-life temperature of papain-inhibitory activity.39G0. C. H.13G0.51G0.2 R8N/T71G ND D15S/C83K ND P13F/G1O8S 70.73G0.4 L9V/C83W ND Cycle 2.37G0.13 ND 1.8 71.47G0.4 Site-Directed Mutagenesis Optimization of glycosylation of human cystatin C was successfully conducted by application of RCG to obtain the best papain inhibitory activity as shown in Table 9.0 V10S/Y102G 59.60G0.60G0.07e ND ND 2. J. d ND.4 ND 66.96G0.28G0.49G0.8 64. b MeanGS.4 ND 68.22 e 3.0 S2K/V57F 68.51G0. random search E20K/S115D 59.5 P6C/Q107L 72.20 1.27G0. LLC .55G0. and Nakai.05.29e 1. D. Dou.03G0. S.0 L27S/AI20R 62.03 0. Scaman.17G0.76G0.22 0.4 E21S/L64G ND R25Y/H90N 65.4 65.06e Variants (one site) T1/2a(8C) Relative Activityb Cycle 1. which is suitable for multifactor optimization.13 1.4 68.3 73.11 0.68G0. random search L9F/C83H 67. 1 2 q 2006 by Taylor & Francis Group.12 0.00G0. Multifactor optimization required for those applications.6 73.18e 1.43e 3.12 1.0 Cycle 1. no data because of absence of detectable papain-inhibitory activity after puriﬁcation process.14e 2.09 0..09e 0.24G0.5 Cycle 2.03e 0. Biochimica et Biophysica Acta.65G0.16e 1..14e 3.57G0.7. 115–124.8 71.8 L9F/R93L 67.3 G4L/D40I 58.06e 2.12e 2.97G0.13 0. Jing. is feasible.3 A16W/R93V 58.97G0.1 72.D.31G0. c WT.85G0.9 Left half and right half are data for double mutants and single mutants.2 F29K/S44R 67. S.01G0.09e ND 0. centroid search D15P/H86I 71.02 2. Source: From Ogawa.30e 4.17G0.EXPERIMENTAL DESIGN AND RESPONSE-SURFACE METHODOLOGY 317 Broader application of RCO to art work or commercial designs may be possible.4 70.35e 4. M.0 72.4. respectively.. e Signiﬁcant differences from WT using LSD analysis at P!0.07e L9F P6C G32Y D15L G4L F29K. 1599.7 Thermostability and Activity of Human Cystatin C Variants (nZ3) Variants (two sites) WTc T1/2a(8C) 68.11G0.5 59.7 D15L/G69I 63.4 66.2 Relative Activityb 1.50G0.08 1. 9.19G0.01e ND 0. wild-type.20e 2. D.98G0.3 59. Theoretically.

Nakamura.. Figure 9. M. Biochimica et Biophysica Acta.318 HANDBOOK OF FOOD AND BIOPROCESS MODELING TECHNIQUES Figure 9.. J.8d demonstrate a decrease in helix (increase in the scale) and strand. H. The reaction mechanism of unexpected amyloidosis as suggested by maps is an advantage of RCG.8b match the active site and binding site. 2002.. C. (c): a-helix propensity. 1599.24 An approximate ﬁve-fold enhancement of activity was obtained by mutant G12W/H86V compared to the recombinant control of wild-type enzyme. Figure 9.. S.. Kitts. respectively. 115–124. q 2006 by Taylor & Francis Group. H.8e shows that a decrease in bulkiness is favorable for the activity. Scaman. which was performed without prior information on the amyloidosis. and Nakai. D.. (From Ogawa.. LLC . (a) and (b): different domains in the sequence. D. (e): bulkiness. S. (d): b-strand propensity. Jing. Dou.8 RCG maps of site-directed mutagenesis of human cystatin C. Figure 9.) avoid amyloidosis (insolubilization) during mutation and puriﬁcation from Pichia pastoris cells. with the substrate papain.8a and Figure 9. which may lead to a decrease in amyloidosis.8c and Figure 9.

a user-friendly interface is not available to inexperienced users. This approach may be a good strategy in enhancing both the objectivity and the reliability for the global optimization. such as simulated annealing and genetic algorithms. multiple regression analysis (MRS) has been widely used as in the case of a variety of RSM. but it may nonetheless be impossible. such as Statisticae. multicolinearity may hinder accurate prediction of objective functions by MRS. Recent applications of RSM in food/bio processing are shown in Table 9. especially in the case of automated optimization using computers. RCO Despite the advantages of RCO over SSO as discussed above.EXPERIMENTAL DESIGN AND RESPONSE-SURFACE METHODOLOGY 319 Table 9.8 SSO vs. usually include a sophisticated graphic subroutine to depict predicted response surfaces.28 However.8.36 q 2006 by Taylor & Francis Group.25 This problem is best circumvented by relating the objective response variables to the principal components of the predictor variables. This process can be replaced by the SSO after adequately narrowing down the search spaces by bringing them to the neighborhood of the global optimum by ﬁrst using RCO. random EVOP.4. automated optimization of many analyses or processes could be performed. principal components thus obtained were related to objective functions using artiﬁcial neural networks (ANNs). Typical examples are Box EVOP. Use of RCO to do the same is difﬁcult due to required human intervention in the response-surface mapping. the manual simultaneous shift in the RCO program was used as a ﬁne adjustment of search direction toward the global optimum. By using SSO as a subroutine.26 The ANN packages. 9.4. To analyze structure-activity relationships. There have been many attempts at applying the computerized optimization. Optimization of GC patterns to obtain the best wine blending27 and blending of GC fractions with concentrated fruit juice to best simulate fresh juice were realized. and simplex EVOP.7 Multivariate Response Surfaces For depicting response surface. the SSO could be advantageous.8 Recent Applications of Response-Surface Methodology in Food/Bio Processing Method Central composite designs Food/Bio Processes High pressure milk coagulation Pork-batter gel Gelled dairy deserts High-melting milk fat Wine blending Finding Best gel strength Best texture Best texture Higher melting characteristics Best blend based on GC proﬁles Better formulation than mixture designs Best taste Best loaf volume Reference 30 31 32 33 27 22 34 35 Mixture designs Simplex optimization Random-centroid optimization Food formulation Cooked Indica rice Dietary-ﬁber bread 9.29 As was previously mentioned. LLC . rotating square EVOP. However. GLOSSARY Evolutionary operation (EVOP) A group of algorithms to shift or rotate the initial design pattern to move toward the optimum.

and Colosimo. Lagrangian multiplier Supplemental/hypothetical independent variables when the number of experiments is less than the number of independent variables in linear regression analysis. Weinheim: Wiley-VCH. Taguchi. 2nd ed.. A. 1996. 1471–1491. C. 891–901. Zbilut. 5. and then factorial (quadratic) multiple linear regression analysis to compute the correlation of factors with responses. R. L.17 Taguchi’s orthogonal designs One of the most efﬁcient fractional factorial designs proposed by Genichi Taguchi.. Centroid The original simplex optimization technique deﬁnes centroid as the average point of nK1 vertices after excluding the worst vertex (Figure 9. and Montgomery. M. 1987. Spendley’s matrix Design of the starting simplex vertices as suggested by Spendley. Full factorial designs Effects of not only main factors but also interactions between all possible combinations of factors.. Stevenson. 2002. A. A. RCO for genetic engineering (RCG) RCO to apply for site-directed mutagenesis of peptides and proteins. Mermet. C. therefore. R. 2nd ed. including two-factor. New York: Marcel Dekker. Central composite designs Typical designs for the original response surface methodology as shown in Figure 9. International Immunopharmacology.. q 2006 by Taylor & Francis Group. J. H.M. Simplex Pinnacles of vertices in SSO. 2. Mixture designs In food formulation... Benigni.. Otto.320 HANDBOOK OF FOOD AND BIOPROCESS MODELING TECHNIQUES Fractional factorial designs In addition to computing effects of main factors..6). 2002. L. M. 1. M. 3. A similar deﬁnition is used in RCO. 2001. Webber. Giuliani. However. this equation should be included in multiple regression equations as a constraint. Otto.. J. System of Experimental Design: Engineering Methods to Optimize Quality and Minimize Costs. Chemical Reviews.. Khuri. and Cornell. P. This is a typical case requiring a Lagrangian multiplier. M.. I. Myers. Response Surfaces: Designs and Analyses. Widmer. P. S. Anti-allergic properties of fermented foods: an important immuno-regulatory mechanism of lactic acid bacteria?. Eds. Plackett–Burman design This fractional factorial design aims to deﬁne the effects of main factors only3 without computing the effects of interactions.. and J. Kellner. R.-M. G.. L. 4. only selected interactions between factors are considered to reduce the total number of experiments to be carried out. or even higher-level interactions are investigated. 102. ingredients are used as the independent variables in a similar form to the full factorial designs. and Gill. LLC . REFERENCES 1. 1998. pp. 759–773. Sequential simplex optimization (SSO) Spendley’s matrix is used as a starting simplex that is then iteratively moved in different directions depending on larger (maximization) or smaller (minimization) response values. New York: Wiley. D.. New York: Unipub. In Analytical Chemistry. Random-centroid optimization (RCO) A regulated random design assisted by surface-response maps as human intervention effective for searching the global optimum. Cross. Sirabella. Response-surface methodology (RSM) The original RSM use central composite designs for experiments. Response Surface Methodology: Process and Product Optimization Using Designed Experiments. the total contents of ingredients should be 100%. 6. Optimization and experimental design. A. three-factor.3. J... Nonlinear signal analysis methods in the elucidation of protein sequence-structure relationships. H.

and Deming. Statistics for Experimenters: An Introduction to Design. Nakai.. M. Cornell.. Parker. 52. 449–457. An attempt to minimize potassium sorbate concentration in sponge cake by modiﬁed atmosphere packaging combination to prevent fungal spoilage.. L. 29. ´ s.. 2000. Alizadeh-Pasdar. Datta. 1990.. Design of Experiments. 4. Dou. E. An investigation of genetic algorithms for the optimization of multi-objective ﬁsheries bioeconomic models. R..... S. H. Ross.. E. Y.. Taguchi Techniques for Quality Engineering. 11. K. 15. K. 1991. G. New random-centroid optimization program for windows useful in research and development. M. J. S. Destouesse. Sequential Simplex Optimization. D. E. 1642–1654. 2003.. C. 2002.. J. Blattner... Journal of Agricultural and Food Chemistry. 1957.. W. Dou. International Journal of Production Economy. J. and Nakai. R. 2000.. 2770–2775.. 5277–5283. S. Data Analysis. Hicks. Esclapez. 2002.. Arteaga.. Alternative method for the quantiﬁcation by gas chromatography triacylglycerol class analysis of cocoa butter equivalent added to chocolate bars. International Transaction and Operation Research. Experiments with Mixtures: Designs.. S. Indianapolis IN: Que. Effect of macerating enzymes on red wine aroma at laboratory scale: 14. 78. New York: McGraw-Hill. and Nakai.. S. 5515–5523.. B. 243–254.. 20. Moustirats. H. Enhancement of proteinase inhibitor activity of recombinant human cystatin C using random-centroid optimization. S. Hydrophobic Interactions in Food Systems. E. Hext. 21. Victor. Journal of Agricultural and Food Chemistry... and Valle Exogenous addition or expression by transgenic wine yeasts. and Kanagawa. Amantea. J.EXPERIMENTAL DESIGN AND RESPONSE-SURFACE METHODOLOGY 321 7. S. and Nakai. Nakamura. 1599. Pascoe. International Journal of Food Property. 17.. Homology similarity analysis of sequences of lactoferricin and its derivatives. Salagoity. Vol. Guyno. Optimization of site-directed mutagenesis. 1988... Nakai. M. ﬂuorescent probe and solvent accessibility study of egg and milk proteins. 5. 45. S. Chan. 51. M. W.. J. 26. C. and Stewardson. Marin. Boca Raton FL: CRC Press. Sanchis. 2005. q 2006 by Taylor & Francis Group. Y. L. K. 441–461. 1. Boca Raton. 6. 26. and Ramos. S. 25.. and Himsworth. LLC . M. S. S. S. 2001. 16. 49. Li-Chan. 1993. Toma. D. Gil.. C. 1962. 2002. N. 10. Food Microbiology. Systematic experimental designs for product formula optimization. Spendley. 115–124. A. and Hunter. Gueon. and Funane. Ogawa. 8. Box.. 1215–1223. M. Ogawa. Taguchi’s system of experimental design and data analysis: a quality engineering technology for the food industry.. H.. 1994. 2004. and Nakai. 24. 5. 311–322.. New York: Wiley.... 13. J. FT-Raman spectroscopy.. Technometrics. C. V.. Trends in Food Science & Technology. 33–49. S. Nakai.. Morgan. and Medina. S. S. G. Walters. C. 25–47. 9. G.. Nakai. C. FL: CRC Press. G. 1998.. and Tamiz. Excel 2000 Functions in Practice. Dou. Tokyo: Maruzen.. Journal of the Society of Dairy Technology. S. Pongcharoen. 1996. 1.. 2nd ed. 57.. M. 52.. J. S. Charteris. 2003. and the Analysis of Mixture Data.... Deﬁnition of outliers using unsupervised principal component similarity analysis for sensory evaluation of foods. Braiden. D. E. 289–306.. Nakai. 33–49. Scaman.. and Scaman. and Ogawa.. Taguchi.. and Nakai. M.. A. 2004. S. Journal of Agricultural and Food Chemistry. Sequential application of simplex designs in optimization and evolutionary operation. Models. P. 21. F. J. 19. W. New York: Wiley. 27–37. Li-Chan. Kitts. G. R. 178–182. 18. V. Mardle.. J. 46. and Model Building.-H. Journal of Agricultural and Food Chemistry. S. P. Vazquez-Arteaga. M. S. F. C. Biochimica et Biophysica Acta.. Hunter. 27. 22. Nakai. Dou... 2004.. H. 7. Determining optimum genetic algorithm parameters for scheduling the manufacturing and assembly of complex products.. N. S.. J. F. 1992. 28. 23. Journal of Food Science. E. Journal of Agricultural and Food Chemistry. P. Li-Chan. S.... Ogawa. H. and Li-Chan. P. 1978. Dou. A computer-aided strategy for structurefunction relation study of food proteins using unsupervised data mining. Computer-aided optimization of wine blending. P. J. J... E.. M.. Food Research International. S.. C. 12. Random-centroid optimization for food formulation. J. Nakamura. Jing. International Journal of Food Property.

42. T334–T344. 34. Nor Aini. Optimum cooking conditions for Indica type rice by using random-centroid optimization. q 2006 by Taylor & Francis Group. microbial transglutaminase and heating temperature on pork batter gel properties. Verbeken. 43–54. K.. K. Mamot.. Optimum composition in ingredients for highﬁber bread containing wheat bran by random-centroid optimization. International Journal of Food Science and Nutrition.. Some techniques of evolutionary operation. I. and Li-Chan.. 4. Effect of high pressure processing on rennet coagulation properties of milk. 35. 1964. Nishimura.. Y... 53. 31. Journal of Cookery Science Society of Japan. Ramaswamy.. 245–256. N. Innovative Food and Emerging Technologies. 2002. Refer to 16... Pietrasik.. P. LLC .... M. 9–16. W. Y. E. M. 1997. C. and Dewettinck.. 35. and St-Gelais. I.. Transaction of the Institute of Chemical Engineering. K. and Noor Lida. Thas. 18. A.. Food Research International. 32. H. S. 31–37. 2004.. Z. O. 30.. C.. 817–823. Horimoto. S. Nor Hayati. Food Hydrocolloids. and Chiji. Aminah. 2003. Low. Textural properties of gelled dairy desserts containing k-carrageenan and starch. Response surface methodology study on the effects of salt. H. Kobayashi.. S. 387–396. Physical characteristics of modiﬁed milk fat in high-melting fat preparation. Imazuya. Nakai. S. Goto.322 HANDBOOK OF FOOD AND BIOPROCESS MODELING TECHNIQUES 30. D. 33. and Nakai. Y. Bulletin of Fuji Womens’ College Number 35. 36. 1997. 2002. Pandey. H. Series II. D.

- None 1
- Pobedinsky
- enzime panificatie
- Acizi
- chapter 13
- Capitolul_16
- Capitolul_18
- Novel Food Processing. Effects on Rheological and Functional Properties
- Biochimie
- Legea 121 Admin Si Deetatizare
- Societatea civila si politicile de mediu
- prezentare ape romania
- Legea Privind Finantele Publice Locale Nr.397-XV
- Legea Privind Protectia Consumatorilor
- Legea Privind Finantele Publice Locale Nr.397-XV
- Lege privind descentralizarea administrativă
- Legea ocrotirii sanatatii
- Manualul Calitatii Apele Romane
- Studiu Idis Viitorul
- Ingineria Designului de Produs
- Proiectare produse noi
- MARKERI DE CALITATE ȘI AUTENTICITATE A
- Betaxanthins as Substrates for Tyrosinase. an Approach
- Proteine

- AIAA-2002-5631
- Ebeb2016 Slides
- Doe
- EC2040_T4
- Ventury Scrubber Paper
- CBLDD07
- Nips Demoreport 120709 Research
- M5L1_LN
- Production of Sterilizing Agents from Calendula officinalis Extracts Optimized by Response Surface Methodology
- A multiperiod degree constrained minimal spanning tree problem
- Lecture1-PO SS2011 01 Introduction p26
- PPT- NSGA-II
- Sensityvity Analysis
- p-median model facility location
- tabu_TSP
- Distributed Constraints Slides
- vwf_Nov2012
- 6 Numerical Methods for Constrained Optimization.pdf
- ISI 8- Optimum Design of Multi-span Composite Box Girder Bridges Using Cuckoo Search Algorithm
- 05493959.pdf
- Course Outline
- DAA PPT
- Advances Vol3 No5 Dec2013
- TSP DynamicProgramming
- LAGRANGE MULTIPLIER.pdf
- Infosheet Mf Starccm 2013 Final
- Chalmers University of Technology Master s Thesis (after correction)
- Bv Cvxslides
- Competitive Distribution System Planning Model Integration Of
- A New Solution to the Optimal Power Flow Problem
- Capitolul_9

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue reading from where you left off, or restart the preview.

scribd