MS ue acer lose Bel tt ed OMY A | LEY Statistics for Expermenters Design, Innovation, and DiscoveryHow, with a nininum Son. ean you discouer what does what to what? Uhich factors do what to which responsts? Only tr exceptional circumstances do you need or should you attempt to answer all the Guestions: ath one. experiment, Ta un ongoing mucstigation, a rough rule ts that only a porlion (say 25%) Sf he experimental, (gon and. budget should be invested in he first design. tctions called, /for asa result ar G perinnt are: fluo hin 2 Soy» oy? vey A, , ZL Cushing a on new knowledge, 2 Using the new hnowledge lo book for, efeurther possibilti The business Sf! Ye ts. to endeavor ta find out what. you dour know from. cokat « t Gow dey that's what F called. ‘guessing what was: on the. other side o the hill.” (i Duke Sf Ses ltinytonf* The best tine to plan.an experiment softer you ve done tt. (R.A. Fisher) The. democratization oS Seientife Atethod. ory model isan opproximuatian, Tis the data that avercal. (They uctualy huppenedy The model isa hypothetical confediure that might or might not stammartse andlor eapslain anportant patures Sf ‘the data. Ml models are LOVONGy SOL nodels are agoful, Don rfl tr love with amok. I ts not unusual for a welt-disigned ¢ apecinent lo anily xt ts Te fund ‘out what 4oppens wher you change something, it ts.necessary to change a Ts. better to solve the right problem opproxinatily than to the corong problim enuctly (FU Tibey) & yperiuent and you Useel ( Cole. Porter) * Sh some cases attributed quotes have been added.va nyinect who does not know experimental design WOO GN CY INEEL, (Renak. made to one. She authors by an. ercatlioc oS the oyola Votor Gompuny) The narrow inadequate paradigen fa strenght line: “ from gestion lo anstocr, (Sohn W Fibey) Reifection snot possible: as always an ufproxination. “ Wost often the problem és not tomuke afnal decision but to see wha 3 worth “ying. “Block what LY Ul Cle and randomize tn hel gou cant” can offre. rimately, feestily aR anelysis ‘af standard: assumptions TOMI, “ The laggest member y ang groups ts harye—but wd arcoptionally lagge? Where there are three op four machines, one will be substantially better or worse. than the others. (oll: Ou) That a concleston reached in one cnowonnent fouy the laboratory) will opply ne kfferent notronnunt fsay the full-scale proce but on what Deming called “a kup of fu hnowledge can narrow the chasm but not elaninate tt. rss] bs based not on statistical, PCUSOWING Good statistics and subject matter Discovering the anespected, és more inportant than confirming the known. One must LYy Ly dong the. hing for though, you think, you hnow tt, Got have no certainly eatil, YOu UY. (ophocts) Mnong te factors: lo be consiedered there will wasteally be. the otal few and the trivial my. GE ia A, furan) — flOCSS should be. routinely operated so as ta prodace nol only Product but dormation on how to auprove the product. Joruetines the: only thing p you can do with a poorly designe Sad out what Hdd, ‘of. ‘ (R. ut. Fisher) yperanint ts to lrg to a She tiperinenter who belwoes that only ong factor ata tine should be varied, x amply, provided for ly “sig © factorial axberiment. > 7 Wy 2 . : Com, Gf! the probabiliy ws OM gy no finding «crock I: gold, behind the next tree, wouldn “you go aud look?Designing an experinent ts lhe. eyembling with the deuit- Only @ random slratigg can fiat all his betting systeus. (R.A, Fisher) The most ening plrase to hear in scence, the one that heralds discovirtes, is not “Gurchal” but “Now that's fing...” (Fsaace sinew) ~tayone who has never make a mistake has never tied anything HW, fe bert Sastein) lee eouything that can be counted counts and not cocrything that counts can be: counted. ( Ubert Caster) » - 7 . : Compunsatory adjustment ty essential when, coc though Oe Can correctly asnign the cause of disturbances, woe cannot othiriise eliminate then. és j % y o Sec ny The idea ofa Process aa [Hifect stale oS control contraocnis the Second Law oS Thermodynamics: Thus aan eauct MME Y ‘control is unrealizable, and must be regarded us a purely theoretical, concept. We shoul! not be gfraid, of finding od soucthing. JSechk computer, Programs that allow pgou to do the thinking : When. the. ritio oS ‘he lacgest to the smallest obscroation Y mara» ao licge, Low howld question whether the dita are being analyzed a the right nulttc: (wransformation). Original data should be presented, ma wey that will, PLES the evidence th the original data, (Walter Shewhurt) Lou can see a lot by, fest looking. (Soy? Berra) computer should make both calculations and, ‘graphs. Both sorts Of ouput should be studied: cach wl contribute to undestanding. (FF. brscombe) . Uhuply works hard to: ensure that anything that CAN YO WONG will [GP WONG: With un adequate. syste of prowess montlorig, theryfore, more and more of he. things that can go wrong will bv corrected and more and more oS NMuiphy &s riches wll be permanently stymied. There's newer bein a signal ‘without notse.tre OSC Y and @ i furclor Dy be correlated with cach other ony because hey are each correlated with some “lurking variable” z. When rang an erperduent the sifest assuinption tx hat untss extruordinary Precautions are tahun, it will be run invorrectls. Knowledge os four. (Sir Frans. Bacon, Show we the data! We are the muste muhers, we are the builders of drcums. (O'Shaughnessy) Few things are less common Ut COMNION SERS. Criteria must be reconsidered at cory staye of an inoestigation. Many uxyful time sertes models care be egarded as. reed, hes for transforming serial data into white notes With sequential ‘assembly, dexigns can be bul up sa that the comple rity ihe design mattehes that Sf the ‘problem. . Mhen the. syygnal, “3 oppry priately scaled, does tt look the pare ft he noise? Suralive tnductive-deductioe proble m soloing tingeared to the structure yf the fuinan brain and is are of the cary diy caperince. When You ser the credits coll at the end ofa successfil mavte, You realize there are MANY. things that must be attended to tn addition to choosing a good serppt. dimilerly an PULNENG a succesyfiel apparent, there are MUY things that must be atte mide to tn addition to choosing agood experimental desiye: a ‘ Siva capital ‘mistake to theorize bore ane has data. fe Sherlock Holmes in Scandal tn Bokunia "by Sim ethur Conan Doyle) Gorrelution can supply a@ pute oS Gukages betevecn variables that can be g Wpirant from, and sonctines toully uurclated ta, Pultern for Causation. When Vauply spruks, listen: M factorial design makes cory observation do double (aultiple) duty. (Suck Coutin)WILEY SERIES IN PROBABILITY AND STATISTICS Established by WALTER A, SHEWHART and SAMUEL S. WILKS Editors: David J. Balding, Peter Bloomfield, Noel A. C. Cressie, Nicholas 1. Fisher, Jain M, Johnstone, J. B. Kadane, Louise M. Ryan, David W, Scott, Adrian F, M. Smith, Jozef L. Teugels Editors Emeriti: Vie Barnet, J. Stuart Hunter, David G, Kendall A complete list of the titles in this series appears at the end of this volume.Statistics for Experimenters Design, Innovation, and Discovery Second Edition GEORGE E. P. BOX J. STUART HUNTER WILLIAM G. HUNTER ON RSCI ENCE A JOHN WILEY & SONS, INC., PUBLICATIONCopyright © 2005 by John Wiley & Sons, Inc. All rights reserved, Published by John Wiley & Sons, Inc,, Hoboken, New Jersey. Published simultaneously in Canada. No part of this publication may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, photocopying, recording, scanning, or otherwise, except as permitted under Section 107 of [08 of the 1976 United States Copyright Act, without either the prior written permission of the Publisher, or authorization through payment of the appropriate per-copy fee to the Copyright Clearance Center, Inc,, 222 Rosewood Drive, Danvers, MA 01923, 978-750-8400, fax 978-750-4470, or on the web at www.copyright.com, Requests to the Publisher for permission should be addressed to the Permissions Department, John Wiley & Sons, Inc.. 111 River Suet, Hoboken, NJ 07030, (201) 748-6011, fax (201) 748-6008, or online at hup:/www.wiley.com/go/permission. Limit of LiabilityDisclaimes of Warranty: While the publisher and author have used their best efforts in preparing this book, they make no representations or warranties with respect to the accuracy or completeness of the contents of this book and specifically disclaim any implied warranties of merchantability or fitness for a particular purpose. No warranty may be created or extended by sales representatives or written sales materials, The advice and strategies contained herein may not be suitable for your situation. You should consult with a professional where appropriate. Neither the publisher nor author shall be liable for any loss of profit or any other commercial damages, including but not limited to special, incidental, consequential, or other damages. For general information on our other products and services please contact our Customer Care Department within the U.S. at 877-762-2974, outside the U.S. at 317-572-3993 or fax 317-572-4002. Wiley also publishes its books in a variety of electronic formats, Some content that appears in print, however, may not be available in electronic format. For more information about wiley products, visit our website at www.wiley.com. Library of Congress Cataloging-in-Publication Data: Box, George E. P. Statistics for experimenters : design, discovery, and innovation / George E.P. Box, J. Stuart Hunter, William G. Hunter.- 2nd ed. p. cm. Includes bibliographical references and index. ISBN-13 978-0471-7 1813-0 (acid-free paper) ISBN-10 0-471-71813-0 (acid-free paper) 1, Experimental design 2. Analysis of variance. 3. Mathematical statistics, 1. Hunter, J. Stuart, 1923 U1. Hunter, William Gordon, 1937— IIL-Title. QA279.B69 2005 519,5—de22 2004063780 Printed in the United States of America. 10987654321To the memory of WILLIAM GORDON HUNTER WILLIAM G. HUNTER 1937-1986Experiment! Make it your motto day and night. Experiment, And it will lead you to the light. The apple on the top of the tree Is never too high to achieve, So take an example from Eve... Experiment! Be curious, Though interfering friends may frown. Get furious At each attempt to hold you down. If this advice you only employ, The future can offer you infinite joy And merriment ... Experiment And you'll see! Coxe Porter* *“EXPERIMENT™ from “NYMPH ERRANT,” c. & a. Cole Porter. Copyright © 1933 larms Inc. Reproduced by kind permission of CHAPPELL & COMPANY LIMITED for the territory of World (ex. U.! Canada and Scandinavia). © 1933 WARNER BROS. INC, Copyright Renewed. All Rights Reserved. Used by permission.When the Lord created the world and people to live in it—an enterprise which, according to modem science, took a very long time—I could well imagine that He reasoned with Himself as follows: “If 1 make everything predictable, these human beings, whom I have endowed with pretty good brains, will undoubtedly learn to predict everything, and they will thereupon have no motive to do anything at all, because they will recognize that the future is totally determined and cannot be influenced by any human action. On the other hand, if I make everything unpredictable, they will gradually discover that there is no rational basis for any decision whatsoever and, as in the first case, they will thereupon have no motive to do anything at all. Neither scheme would make sense. I must therefore create a mixture of the two. Let some things be predictable and let others be unpredictable. They will then, amongst many other things, have the very important task of finding out which is which.” E. F, ScHuMACHER* *From Small Is Beautiful. Used by permissionGreek Alphabet alpha beta gamma delta epsilon zeta eta theta iota kappa lambda mu Nv ag Oo a Pp Loa Tt Tu oo XxX vy Qo nu xi omicron pi tho sigma tau upsilon phi chi psi omegaContents Preface to the Second Edition Chapter 1 Catalyzing the Generation of Knowledge 1d. 1.2. 1.3. 1.4, 1.5. The Learning Process Important Considerations The Experimenter’s Problem and Statistical Methods A Typical Investigation How to Use Statistical Techniques References and Further Reading Chapter 2 Basics (Probability, Parameters, and Statistics) 2.1. 2.2. 2.3. 2.4, 2.5. 2.6. 2.7, 2.8. 2.9. 2.10. 2.11. 2.12. 2.13. 2.14. Experimental Error Distributions Statistics and Parameters Measures of Location and Spread The Normal Distribution Normal Probability Plots Randomness and Random Variables Covariance and Correlation as Measures of Linear Dependence Student’s t Distribution Estimates of Parameters Random Sampling from a Normal Population The Chi-Square and F Distributions The Binomial Distribution The Poisson Distribution xvAppendix 2A. Mean and Variance of Linear Combinations of Observations References and Further Reading Chapter 3 Comparing Two Entities: Reference Distributions, Tests, and Confidence Intervals 3.1. 3.2, 3.3. 3.4. 3.5. 3.6. 3.7. 3.8. Relevant Reference Sets and Distributions Randomized Paired Comparison Design: Boys’ Shoes Example Blocking and Randomization Reprise: Comparison, Replication, Randomization, and Blocking in Simple Experiments More on Significance Tests Inferences About Data that are Discrete: Binomial Distribution. Inferences about Frequencies (Counts Per Unit): The Poisson Distribution Contingency Tables and Tests of Association Appendix 3A. Comparison of the Robustness of Tests to Compare Two Entities Appendix 3B. Calculation of reference distribution from past data References and Further Reading Chapter 4 Comparing a Number of Entities, Randomized Blocks, and Latin Squares 4.1, 4.2, 43. 44. 4.5. Comparing & Treatments in a Fully Randomized Design Randomized Block Designs A Preliminary Note on Split-Plot Experiments and their Relationship to Randomized Blocks More than one blocking component: Latin Squares Balanced Incomplete Block Designs Appendix 4A. The Rationale for the Graphical ANOVA Appendix 4B. Some Useful Latin Square, Graeco—Latin Square, and Hyper-Graeco—Latin Square Designs References and Further Reading Chapter 5 Factorial Designs at Two Levels 5.1, Introduction ' CONTENTS s7 67 67 81 92 94 94 105 110 112 117 120 123 133 133 145 156 157 162 166 167 168 173 173CONTENTS 5.14. 5.15. 5.16. 5.17. 5.18. . Example 1: The Effects of Three Factors (Variables) on Clarity of Film . Example 2: The Effects of Three Factors on Three Physical Properties of a Polymer Solution . A.23 Factorial Design: Pilot Plant Investigation . Calculation of Main Effects . Interaction Effects . Genuine Replicate Runs . Interpretation of Results . The Table of Contrasts . Misuse of the ANOVA for 2 Factorial Experiments . Eyeing the Data .. Dealing with More Than One Response: A Pet Food Experiment . A 24 Factorial Design: Process Development Study Analysis Using Normal and Lenth Plots Other Models for Factorial Data Blocking the 2 Factorial Designs Learning by Doing Summary Appendix 5A. Blocking Larger Factorial Designs Appendix SB. Partial Confounding References and Further Reading Chapter 6 Fractional Factorial Designs 6.1. 6.2. 6.3. 6.4. 6.5. 6.6. 6.7, 6.8. 6.9. 6.10. 6.11. Effects of Five Factors on Six Properties of Films in Eight Runs Stability of New Product, Four Factors in Eight Runs, a 24~! Design A Half-Fraction Example: The Modification of a Bearing The Anatomy of the Half Fraction The 27; Design: A Bicycle Example Eight-Run Designs Using Table 6.6: An Illustration Sign Switching, Foldover, and Sequential Assembly An Investigation Using Multiple-Column Foldover Increasing Design Resolution from III to IV by Foldover Sixteen-Run Designs 174 175 177 178 181 183 185 186 188 190 193 199 203 208 211 215 219 219 221 222 235 235 236 239 240 244 246 247 249 252 257 258xii 6.12. 6.13. 6.14. 6.15. 6.16. The Nodal Half Replicate of the 2° Factorial: Reactor Example The 2{7* Nodal Sixteenth Fraction of a 2 Factorial The 2}3-'' Nodal Design: The Sixty-Fourth Fraction of the 2'> Factorial Constructing Other Two-Level Fractions Elimination of Block Effects References and Further Reading Chapter 7 Additional Fractionals and Analysis 7A. 7.2. 73. Plackett and Burman Designs Choosing Follow-Up Runs Justifications for the Use of Fractionals Appendix 7A. Technical Details Appendix 7B. An Approximate Partial Analysis for PB Designs Appendix 7C. Hall’s Orthogonal Designs References and Further Reading Chapter 8 Factorial Designs and Data Transformation 8.1. 8.2. A Two-Way (Factorial) Design Simplification and Increased Sensitivity from Transformation Appendix 8A. Rationale for Data Transformation Appendix 8B. Bartlett's x for Testing Inhomogeneity of Variance References and Further Reading Chapter 9 Multiple Sources of Variation 9.1. 9.2. 9.3. 9.4. Split-Plot Designs, Variance Components, and Error Transmission Split-Plot Designs Estimating Variance Components Transmission of Error References and Further Reading Chapter 10 Least Squares and Why We Need Designed Experiments 10.1. 10.2. 10.3. Estimation With Least Squares The Versatility of Least Squares The Origins of Experimental Design CONTENTS 317 317 320 329 329 329 335 335 335 345 353 359 363 364 378 397CONTENTS 10.4. Nonlinear Models Appendix 10A. Vector Representation of Statistical Concepts Appendix 10B. Matrix Version of Least Squares Appendix 10C. Analysis of Factorials, Botched and Otherwise Appendix 10D. Unweighted and Weighted Least Squares References and Further Reading Chapter 11 Modeling, Geometry, and Experimental Design Chapter 12 Chapter 13 IL... 11.2. 11.3. 11.4, 11.5. 11.6. Some Empirical Models Some Experimental Designs and the Design Information Function Is the Surface Sufficiently Well Estimated? Sequential Design Strategy Canonical Analysis Box—Behnken Designs References and Further Reading Some Applications of Response Surface Methods 12.1. 12.2. 12.3. 12.4, 12.5. 12.6. Iterative Experimentation To Improve a Product Design Simplification of a Response Function by Data Transformation Detecting and Exploiting Active and Inactive Factor Spaces for Multiple-Response Data Exploring Canonical Factor Spaces From Empiricism to Mechanism Uses of RSM Appendix 12A. Average Variance of Appendix 12B. References and Further Reading Designing Robust Products and Processes: An Introduction 13.1, 13.2, Environmental Robustness Robustness To Component Variation Appendix 13A. A Mathematical Formulation for Environmental Robustness Appendix 13B, Choice of Criteria References and Further Reading xiii 407 410 416 418 420 424 437 441 447 453 454 461 475 483 489 489 503 509 513 518 526 526 528 530 539 539 549 556 558 559xiv Chapter 14 Process Control, Forecasting, and Time Series: An Introduction 14.1. Process Monitoring 14.2, The Exponentially Weighted Moving Average 14.3. The CuSum Chart 14.4. Process Adjustment 14.5. A Brief Look At Some Time Series Models and Applications 14.6. Using a Model to Make a Forecast 14.7. Intervention Analysis: A Los Angeles Air Pollution Example References and Further Reading Chapter 15 Evolutionary Process Operation 15.1. More than One Factor 15.2. Multiple Responses 15.3. The Evolutionary Process Operation Committee References and Further Reading Appendix Tables Author Index Subject Index CONTENTS 565 565 569 574 576 585 588 593 595 599 602 606 608 61L 621 623Preface to the Second Edition In rewriting this book, we have deeply felt the loss of our dear friend and col- league Bill Hunter (William G. Hunter, 1937-1986). We miss his ideas, his counsel, and his cheerful encouragement. We believe, however, that his spirit has been with us and that he would be pleased with what we have done. The objectives for this revised edition of Statistics for Experimenters remain the same as those for the first: 1. To make available to experimenters scientific and statistical tools that can greatly catalyze innovation, problem solving, and discovery. 2. To illustrate how these tools may be used by: and with subject matter specialists as their investigations proceed, Developments that would have delighted Bill are the receptive atmosphere these techniques now encounter in industry and the present universal availability of very fast computers, allowing* where necessary, the ready use of computa- tionally intensive methods. Under such banners as “Six Sigma,” management has realized the impor- tance of training its work forces in the arts of economic investigation. With this democratization of the scientific method, many more people are being found with creative ability and unrealized aptitude for problem solving and discovery. Also, the “team idea” not only accelerates improvement but identifies such natural lead- ers of innovation and can allow them to lead. To make such initiatives possible the modern philosophy and methods of process improvement must be taught at all levels of an organization. We believe both trainers and trainees engaged in such efforts will find this book helpful. Also based on a long experience, we *All the computations in this book can be done with the statistical language R (R Development Core Team, 2004), available at CRAN (http://cran.R-project.org). Functions for displaying anova and lambda plots, for Bayesian screening and model building are included in the BHH2 and BsMD R- Packages and available at CRAN under contributed packages. There is as well commercial software, Such as the SCA Statistical System, which some readers will find easier to use. xvxvi PREFACE TO THE SECOND EDITION believe the material in this book provides appropriate training for scientists and engineers at universities whose needs have in the past been frequently neglected. The material of the original book has been rearranged and largely rewritten with the object of ensuring even greater accessibility to its users. In addition, a number of new things have been introduced or emphasized: The many issues, in addition to statistical considerations, that must be attended to in a successful investigation. The néed to work closely with subject matter specialists. The importance of choosing a relevant reference set. The value of graphical techniques as an adjunct to more formal methods. The use of inactive factor and canonical spaces to solve multiple response problems. The greatly increased efficiency sometimes achievable by data transformation using “lambda plots.” The understanding of error transmission and its use in designing products and processess that are least affected by (robust to) variation in system components. The value of split plot and similar arrangements in industrial experimentation particularly in the design of environmentally robust products and processes. The value of a sequential approach to problem solving and in particular the sequential assembly of experimental designs. How to choose the best follow-up runs. The acquiring of investigational technique by hands-on design of some simple device (e.g., a paper helicopter). How empiricism can lead to mechanism. The use of randomization and blocking to allow analysis “as if? standard assumptions were true. The use of complex experimental arrangements, in particular Plackett Burman designs and their analysis. The design information function. An introduction to process control, to forecasting, and to time series analysis. A fuller discussion of evolutionary process operation. USING THE QUESTIONS AND PROBLEMS The questions and problems at the end of each chapter can be used in two ways. You can consider them for review before reading the chapter to help identify key points and to guide your reading and you can consider them for practice and exercise after reading this chapter. You may also find helpful the collection of quotes on the inside covers of this book.PREFACE TO THE SECOND EDITION xvii As you apply these ideas and especially if you meet with unusual success or failure, we shall be interested to learn of your experiences. We have tried to write a book that is useful and clear. If you have any suggestions as to how it can be improved, please write us. ACKNOWLEDGMENTS For help with the preparation of this book we are especially grateful for the generous assistance of Emesto Barrios Zamudio, We also wish to thank René Valverde Ventuara, Carla Vivacqua and Carmen Paniagua Quifiones for their help. We are indebted to David Bacon Mac Berthouex, Sgren Bisgaard, Bob DeBaun, David Steinberg, Dan Meyer, Bill Hill, Merv Muller and Daniel Pefia for their critiques and valuable suggestions on earlier versions of the manuscript and to Claire Box and Tady Hunter who have provided patient encouragement. Throughout this long endeavor we are particularly fortunate in our editor Lisa Van Horn for cheerful and wise counsel throughout the book's preparation. GeorGe E, P. Box J. Stuart Hunter Madison, Wisconsin Princeton, New Jersey April 2005CHAPTERI Catalyzing the Generation of Knowledge 1.1. THE LEARNING PROCESS Knowledge is power. It is the key to innovation and profit. But the getting of new knowledge can be complex, time consuming, and costly. To be successful in such an enterprise, you must learn about learning. Such a concept is not esoteric. It is the key to idea generation, to process improvement, to the development of new and robust products and processes. By using this book you can greatly sim- plify and accelerate the. generation, testing, and development of new ideas. You will find that statistical methods and particularly experimental design catalyze scientific method and greatly increase its efficiency. Learning is advanced by the iteration illustrated in Figure 1.1. An initial idea (or model or hypothesis or theory or conjecture) leads by a process of deduction to certain necessary consequences that may be compared with data. When con- sequences and data fail to agree, the discrepancy can lead, by a process called induction, to modification of the model. A second cycle in the iteration may thus be initiated. The consequences of the modified model are worked out and again compared with data (old or newly acquired), which in turn can lead to further modification and gain of knowledge. The data acquiring process may be scientific experimentation, but it could be a walk to the library or a browse on the Internet. Inductive-Deductive Learning: An Everyday Experience The iterative inductive—deductive process, which is geared to the structure of the human brain and has been known since the time of Aristotle, is part of one's everyday experience. For example, a chemical engineer Peter Minerex* parks “Can you guess why he’s called Peter Minerex? —______ Statistics for Experimenters, Second Edition, By G. E. P. Box, J. S. Hunter, and W. G. Hunter Copyright © 2005 John Wiley & Sons, Inc.2 1 CATALYZING THE GENERATION OF KNOWLEDGE Data (facts, phenomena) cca Deduction Induction Deduction Induction cae Idea (model, hypothesis, theory, conjecture) Figure 1.1. Iterative learning process. his car every morning in an allocated parking space. One afternoon after leaving work he is led to follow the following deductive—inductive learning sequence: Model: Today is like every day. Deduction: My car will be in its parking place. Data: It isn’t. Induction: Someone must have taken it. Model: My car has been stolen. Deduction: My car will not be in the parking lot Data: No. It’s over there! Induction: Someone took it and brought it back. Model: A thief took it and brought it back. Deduction: My car will have been broken into. Data: It’s unharmed and unlocked. Induction: Someone who had a key took it. Model: My wife used my car. Deduction: She probably left a note. Data: Yes. Here it is. Suppose you want to solve a particular problem and initial speculation pro- duces some relevant idea. You will then seek data to further support or refute this theory. This could consist of some of the following: a search of your files and of the Web, a walk to the library, a brainstorming meeting with co-workers and executives, passive observation of a process, or active experimentation. In any case, the facts and data gathered sometimes confirm your conjecture, in which case you may have solved your problem. Often, however, it appears that your initial idea is only partly right or perhaps totally wrong. In the latter two cases, the difference between deduction and actuality causes you to keep digging. This can point to a modified or totally different idea and to the reanalysis of your present data or to the generation of new data.