Teaching the Taguchi method to industrial engineers

Jiju Antony and Frenie Jiju Antony

Introduction
Dr Genichi Taguchi is a Japanese quality management consultant who has developed and promoted a philosophy and methodology for continuous quality improvement in products and processes. Within this philosophy, Taguchi shows how the statistical design of experiments (SDOE or DOE) can help industrial engineers design and manufacture products that are both of high quality and low cost. His approach is primarily focused on eliminating the causes of poor quality and on making product performance insensitive to variation. DOE is a powerful statistical technique for determining the optimal factor settings of a process and thereby achieving improved process performance, reduced process variability and improved manufacturability of products and processes. Taguchi (1986) advocates the use of orthogonal array designs to assign the factors chosen for the experiment. The most commonly used orthogonal array designs are L8 (i.e. eight experimental trials), L16 and L18. The power of the Taguchi method is that it integrates statistical methods into the engineering process. Bendell et al. (1989) and Rowlands et al. (2000) report success of the Taguchi method in the automotive, plastics, semiconductors, metal fabrication and foundry industries. However, Antony (1996) suggests that the application of the Taguchi method in the UK manufacturing and service industries is limited and often applied incorrectly. Moreover, a typical remark is I can do the text book and class room examples, but I am not comfortable and confident in applying the concepts and principles of DOE in my work area. According to Antony et al. (Antony et al. 1996a; 1998a; 1998b; 1999; Antony, 1998) the following issues are key to this lack of or improper application of experimental design techniques based on the Taguchi method: The word statistics invokes fear in many industrial engineers. Many engineers in the UK leave universities without a complete understanding of the power of statistics and are therefore likely to avoid the use of statistical techniques in their subsequent careers. Few graduating engineers have been exposed to applied statistical quality techniques such as DOE, robust design,

The authors Jiju Antony is at the International Manufacturing Centre, Department of Engineering, University of Warwick, Coventry, UK. Frenie Jiju Antony is at the School of Management Studies, Cochin University of Science and Technology, Kerala, India. Keywords Taguchi methods, Statistical process control, Design of experiments Abstract The Taguchi method (Tm) is a powerful problem solving technique for improving process performance, yield and productivity. It reduces scrap rates, rework costs and manufacturing costs due to excessive variability in processes. However, its application by industrial engineers in the UK is limited, in part due to the inadequate statistical education of engineers. This paper presents a simple experiment which can be used in the classroom to teach engineers the basics of the technique and illustrates simple analytical and graphical tools which promote rapid understanding of the results of the experiment. Electronic access The research register for this journal is available at http://www.mcbup.com/research_registers The current issue and full text archive of this journal is available at http://www.emerald-library.com/ft

Work Study Volume 50 . Number 4 . 2001 . pp. 141149 # MCB University Press . ISSN 0043-8022

141

Teaching the Taguchi method to industrial engineers Jiju Antony and Frenie Jiju Antony

Work Study Volume 50 . Number 4 . 2001 . 141149

etc. This is another symptom of the statistical education of the engineering fraternity. Engineers consistently avoid the use of applied statistical techniques in tackling process optimisation and quality control problems. Where techniques are in use, e.g. the use of control charts for process analysis and monitoring, there often appears to be a lack of a full understanding of the basic and fundamental principles behind their application (Morrison, 1997). Many textbooks and courses on DOE primarily focus on the statistical analysis of the problem under study. However, this is but one component of DOE which involves planning, design, execution, analysis and interpretation of results. A lack of communication between the academic and industrial worlds, and between functional specialists restricts the application of the Taguchi method (Tm)and DOE (Antony et al., 1998a). It is important, though too rare, that quality, manufacturing, process, design and operational departments communicate and work effectively with one another.

plastics, automotive, process, metal fabrication, food and electronics and semiconductor sectors (Rowlands et al., 2000).

Typical applications in service industry

The use of Tm in service industries is not often reported. This may be because : service performance is often more difficult to measure; the performance of a service process depends a great deal on the behaviour and attitude of the service provider and it varies with time; and the identification and measurement of control factors and their influence on performance characteristic(s) is often difficult. However, there clearly are possible applications of Tm in the service sector. Examples include: reducing the time taken to respond to customer complaints; reducing errors on service orders; and reducing the length of stay in an emergency room in hospital. If the use of Tm is to become more prevalent, ways must be found to teach engineers (and others) effectively how to apply it successfully.

Potential applications and benefits of using the Taguchi method

The Taguchi method has wide application in manufacturing organisations. Table I illustrates the application of Tm in the
Table I Typical applications of Tm in manufacturing Process/product Injection moulding process Diesel injector Nature of problem

Steps in performing a Taguchi experiment

The process of performing a Taguchi experiment follows a number of distinct steps:

Experiment size

Benefits Annual savings were estimated to be over 40,000 Annual savings were estimated to be over 10,000 Annual savings were estimated to be over 16,000 Process yield was improved by over 10 per cent Biscuit length variability was reduced by over 25 per cent Annual savings were over 30,000

High scrap rate due to 8 trials excessive process variability High rework rate 16 trials

Welding process

Low weld strength

16 trials

Chemical process Biscuit Wire-bonding process

Low process yield Excessive variability in biscuit length Low wire pull strength

8 trials 16 trials 16 trials

142

Teaching the Taguchi method to industrial engineers Jiju Antony and Frenie Jiju Antony

Work Study Volume 50 . Number 4 . 2001 . 141149

Step1: formulation of the problem the success of any experiment is dependent on a full understanding of the nature of the problem. Step 2: identification of the output performance characteristics most relevant to the problem. Step 3: identification of control factors, noise factors and signal factors (if any). Control factors are those which can be controlled under normal production conditions. Noise factors are those which are either too difficult or too expensive to control under normal production conditions. Signal factors are those which affect the mean performance of the process. Step 4: selection of factor levels, possible interactions and the degrees of freedom associated with each factor and the interaction effects. Step 5: design of an appropriate orthogonal array (OA). Step 6: preparation of the experiment. Step 7: running of the experiment with appropriate data collection. Step 8: statistical analysis and interpretation of experimental results. Step 9: undertaking a confirmatory run of the experiment.

Paper helicopter experiment

In many academic institutions within the UK, the focus of engineering statistics is on the theory of probability (for example, card shuffling, dice rolling, etc.), the mathematical aspects of probability and probability distributions (e.g. normal, exponential, binomial, Poisson, log-normal, etc.), hypothesis tests, etc. Quality improvement techniques (DOE, Tm, SPC, etc.) are often not covered. Understandably, graduates are not confident about using such techniques at their place of work. As part of an exercise to increase the awareness of Tm amongst industrial engineers, the authors used a simple paper helicopter experiment, readily used in academic institutions. Due to a limited amount of time, one member from each group in the class was involved with the experimental work. However, the students were all asked to analyse and interpret the data (on an individual basis). The results of

the analysis were discussed in the classroom as part of the process of gaining an understanding of experimental objectives and process. The paper helicopter experiment is quite well known among engineers and statisticians in both the academic and industrial worlds. Many industrial training programmes on Tm use it in some form. However they often focus on the design and analysis of the experiment without providing guidance to engineers on the interpretation of results from the analysis. Moreover, many courses do not cover the importance of careful experimental planning for the success of any industrially designed experiment. The purpose of this experiment was to provide undergraduate engineering students with an understanding of the role of Taguchis parameter design (sometimes called robust design) in tackling both product and process quality-related problems in real-life situations. Parameter design is a well established methodology for improving product and process quality at minimal cost by reducing the effect of undesirable external influences which cause variation in product or process performance (Phadke, 1989). The objective of the exercise was to identify the optimal settings of control factors which would maximise the flight time of paper helicopters (with minimum variation). Here control factors refer to those which can be easily controlled and varied by the designer or operator in normal production conditions. A brainstorming session by a group of students identified six control factors which were thought to influence the time of flight (refer to Table II). Brainstorming should be considered an integral part of the Taguchi methodology it is a useful technique in identifying the most influential factors in an experiment. In order to simplify the experiment, each factor was studied at two levels. The level of a factor here refers to the specified value of
Table II Control factors and their range of settings for the experiment Control factor Paper type Body length Wing length Body width Number of clips Wing shape Labels A B C D E F Level 1 Regular 8cm 8cm 2cm 1 Flat Level 2 Bond 12cm 12cm 3cm 2 Angled

143

Teaching the Taguchi method to industrial engineers Jiju Antony and Frenie Jiju Antony

Work Study Volume 50 . Number 4 . 2001 . 141149

a setting. For example, in the experiment, body width was studied at 2cm and 3cm. Factors at three (and higher) levels make analysis more complicated and are therefore not used in awareness-raising sessions. Having identified the control factors, it is important to list the interactions which are to be studied for the experiment. Interaction exists when the effect of one factor is not the same at different levels of the other factor. An effect refers to the change in response due to the change in level of a factor (Antony et al., 1998b). Consider, for example, the factors wing length and body length of the paper helicopter. Assume each factor was kept at two-levels for the study. Time of flight is the response (or quality characteristic) of interest. Interaction between wing length and body length exists when the effect of wing length on time of flight at two different levels of body length is different. For this experiment, three interactions were identified (from the brainstorming session) as being of interest: (1) body length wing length (B C or BC); (2) body length body width (B D or BD); and (3) paper type body length (A B or AB). The following noise factors were identified (as having some impact on the flight time but being difficult to control): operator-to-operator variation; draughts; reaction time; and ground surface.

Figure 1 Template for paper helicopter design

itself taking about 90 minutes). In this case, the statistical analysis was executed as a homework assignment though the results were discussed in the classroom in detail.

Choice of orthogonal array design

The choice of a suitable orthogonal array (OA) design is critical for the success of an experiment and depends on the total degrees of freedom required to study the main and interaction effects, the goal of the experiment, resources and budget available and time constraints. Orthogonal arrays allow one to compute the main and interaction effects via a minimum number of experimental trials (Ross, 1988). Degrees of freedom refers to the number of fair and independent comparisons that can be made from a set of observations. In the context of SDOE, the number of degrees of freedom is one less than the number of levels associated with the factor. In other words, the number of degrees of freedom associated with a factor at p-levels is (p-1). As the number of degrees of freedom associated with a factor at two levels is unity, in the present example the number of degrees of freedom for studying the six main effects is equal to six. The number of degrees of freedom associated with an interaction is the product of the number of degrees of freedom associated with each main effect involved in the interaction (Antony, 1998). In this simple case, the number of degrees of freedom for studying the three interaction effects is equal to three. Therefore the total degrees of freedom is equal to nine (i.e. 6 + 3). It is important to notice that the number of

One aim was to determine the control factor settings which would best dampen the effect of these noise factors. According to Taguchi, there is an optimal combination of factor settings which counters the effects of noise. In order to minimise the effect of these noise factors, the same student was responsible for all timings reducing the effects of variable reaction times when hitting the stopwatch upon release of the helicopter and its hitting the ground. Figure 1 illustrates a template for the model of a paper helicopter which can be made from an A4 size paper. It forms the basis of a simple experiment, requiring only simple items such as paper, scissors and paper clips. It takes about six hours to design the experiment, collect the data and then perform the statistical analysis (with the experiment

144

Teaching the Taguchi method to industrial engineers Jiju Antony and Frenie Jiju Antony

Work Study Volume 50 . Number 4 . 2001 . 141149

experimental trials must be greater than the total degrees of freedom required for studying the effects. The standard OAs for factors with two levels are L4 L8 , L1 6 , L32 and so on. Here the notation L implies that the information is based on the Latin square arrangement of factors. A Latin square arrangement is a square matrix arrangement of factors with separable factor effects. Here the numbers 4, 8, 12, 16, etc. denote the number of experimental trials. For the helicopter experiment, as the total degrees of freedom is equal to nine, the closest number of experimental trials that can be employed for the experiment is 16 (i.e. L1 6 OA). Having identified the most suitable OA, the next step was to assign the main and interaction effects to various columns of the array. A standard L16 OA (see Appendix) contains 15 columns for either studying 15 main effects, or a combination of main and interaction effects so that the degrees of freedom will add up to 15. In the present example, there are only six main and three interaction effects. This means that only nine columns out of 15 are used. For example, factor D (refer to Table III) was assigned to column 1 and factor C to column 2. Column 3 is empty (see Table III) as the interaction between these factors was of no interest in this experiment. Using the standard linear graphs and OA (Ross, 1988), the remaining factors and interactions were assigned to the columns of an L1 6 in the following manner.
Table III Experimental layout Column no. Factors/interactions Trial no. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 1 D 2 C 4 B 5 BD

Column 1 body width (D), column 2 wing length (C), column 4 body length (B), column 5 body width body length (B D), column 6 wing length body length (B C), column 7 wing shape (F), column 8 paper type (A), column 12 body length paper type (AB) and column 14 number of clips (E). The experimental layout showing all the factors and interactions along with the flight times (measured in seconds) is shown in Table III. As each factor was studied at two levels, coded level 1 represents the low level of a factor setting and level 2 represents the high level setting. Each experiment was replicated in order to capture variation in results due to uncontrolled noise.

Statistical analysis and interpretation of results

In Taguchis parameter design, the basic objective is to identify the conditions which optimise process/product performance. In arriving at this optimal set of conditions, Taguchi advocates the use of signal-to-noise ratio (SNR) the need is to maximise the performance of a system or product by minimising the effect of noise while maximising the mean performance. The SNR is treated as a response (output) of the experiment, which is a measure of variation when uncontrolled noise factors are present in

6 BC

7 F

8 A

12 AB

14 E

Flight time

1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2

1 1 1 1 2 2 2 2 1 1 1 1 2 2 2 2

1 1 2 2 1 1 2 2 1 1 2 2 1 1 2 2

1 1 2 2 1 1 2 2 2 2 1 1 2 2 1 1

1 1 2 2 2 2 1 1 1 1 2 2 2 2 1 1

1 1 2 2 2 2 1 1 2 2 1 1 1 1 2 2

1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2

1 2 2 1 1 2 2 1 1 2 2 1 1 2 2 1

1 2 2 1 2 1 1 2 1 2 2 1 2 1 1 2

2.76, 2.20, 1.93, 2.19, 2.40, 2.82, 3.39, 2.62, 2.46, 2.08, 2.14, 2.05, 2.96, 2.47, 2.62, 2.32,

2.83 2,13 2.30 2.10 2.50 2.31 3.01 2.39 2.12 1.90 2.29 2.12 2.70 2.60 2.91 2.41

145

Teaching the Taguchi method to industrial engineers Jiju Antony and Frenie Jiju Antony

Work Study Volume 50 . Number 4 . 2001 . 141149

the system (Antony et al., 1999). Taguchi has developed and defined over 60 different SNRs for engineering applications of parameter design. For the present study, as the objective was to maximise time of flight it was decided to select the SNR related to larger-the-better (LTB) quality characteristics. This is generally used for quality characteristics such as strength, fuel efficiency, process yield, life of a component and so on. For LTB quality characteristics, the SNR is given by the following equation: 1 1 SNR 10log 1 n y2 i

The other main and interaction effects were calculated in a similar manner (see Table V). Having obtained the average SNR values, the next step is the identification of significant main and interaction effects which influence the SNR. To achieve this, a powerful graphical tool called half-normal probability plots (HNPP) is useful. A half-normal probability plot (HNPP) is obtained by plotting the absolute values of the effects (both main and/or interaction effects) along the X-axis and the per cent probability along the Y-axis. The per cent probability can be obtained by using the following equation: Pi i 0:5 100 n 2

where n = number of values at each trial condition (i.e., 2 from Table II) and yi = each observed value. Table IV illustrates the SNR values (based on equation 1) corresponding to each trial condition. Table V illustrates the average SNR values (SNR) at low (level 1) and high (level 2) levels and the effect of each main and interaction effect on the SNR. Sample calculation for factor C: Average SNR at level 1 of factor C = SNRC 2 = 1/8 * [8.93 + 6.71 + 6.41 + 6.62 +7.12 + 5.95 + 6.89 + 6.38] = 6.88 Similarly, average SNR at level 2 of factor C = SNRC 2 = 1/8 * [7.78 + 8.05 + 10.06 + 7.95 + 9.01 + 8.07 + 8.80 + 7.47] = 8.40 Effect = SNRC 2 - SNRC 1 = 8.40 - 6.88 = 1.52

Table IV SNR table Trial number 1 2 3 4 5 6 7 8 SNR 8.93 6.71 6.41 6.62 7.78 8.05 10.06 7.95 Trial number 9 10 11 12 13 14 15 16 SNR 7.12 5.95 6.89 6.38 9.01 8.07 8.80 7.47

where: n = number of estimated effects (n = 15) and i is the rank of the estimated effect when arranged in the ascending order of magnitude (e.g. for factor C, i = 15). Figure 2 illustrates the HNPP of the factor and interaction effects for the helicopter experiment. The computer software package Design-ease was used to construct the plot. Those effects which are active and real will fall off the straight line, whereas the inactive and insignificant effects will fall along the straight line (Daniel, 1959). The figure reveals that main effects A, C, E and F are statistically significant. i.e. paper type, wing length, number of clips and wing shape are statistically significant. In order to support and justify this claim, another graphical tool (main effects plot) is used. This shows the average SNR values at low and high level settings of each factor. Figure 3 illustrates the main effects plot for the paper helicopter experiment (using the values from Table V). This graphical aid provides non-statisticians with a better picture of the importance of the effects of the chosen control factors. The slope of the line is an indication of the importance of a main or interaction effect. The figure shows that the most dominant factor is the wing length, followed by paper type, wing shape and number of clips. As each factor was chosen at two levels, the effect of

Table V Average SNR table Factors or interactions SNR1 SNR2 Effect estimate D 7.81 7.46 0.35 C 6.88 8.40 1.52 B 7.70 7.57 0.13 BD 7.63 7.65 0.02 BC 7.87 7.40 0.47 F 8.00 7.27 0.73 A 8.12 7.15 0.97 AB 7.66 7.62 0.04 E 8.00 7.28 0.72

146

Teaching the Taguchi method to industrial engineers Jiju Antony and Frenie Jiju Antony

Work Study Volume 50 . Number 4 . 2001 . 141149

Figure 2 Half-normal plot of effects

An interaction plot is useful in providing a rapid understanding of the nature of interactions (Schmidt and Launsby, 1992). Interaction plots are constructed by plotting the average response values (in this case, SNR values) at each factor level combination. Parallel lines are an indication of the absence of interaction between the factors, whereas non-parallel lines are an indication of the presence of interaction between the factors. Figure 4 shows that the effect of body width on the flight time at both levels of body length is the same. In other words, the effect of body width on the flight time is the same, irrespective of the level of body length. This implies the absence of interaction between these two factors.

Determination of the optimal control factor settings

The selection of optimal settings depends on the objective of the experiment or the nature of the problem under study. For the helicopter example, the objective was to maximise the flight time. In Taguchi experiments, the objective is to identify the factor settings which yield the highest SNR these settings will generally produce a consistent and reliable product. Moreover, the process which produces the product will
Table VI Average SNR values Body length 1 1 2 2 Body width 1 2 1 2 Average SNR 7.87 7.54 7.76 7.39

Figure 3 Main effects plot of the control factors

each factor must be assumed to be linear. If non-linear effects are to be studied, it is necessary to choose more than two levels for each factor. However, it is good practice to start off an experiment with two levels and then perform smaller sequential experiments at higher levels to gain a better understanding of the nature of the process. For this experiment, none of the interaction effects is significant. Consider, for example, the interaction between the body length and body width. In order to compute this interaction, the first step is to compute the average SNR values at each of the four combinations of the factor levels. Table VI shows the average SNR values for these four combinations.

Figure 4 Interaction plot between body length and body width

147

Teaching the Taguchi method to industrial engineers Jiju Antony and Frenie Jiju Antony

Work Study Volume 50 . Number 4 . 2001 . 141149

be insensitive to various sources of uncontrollable variation. For the paper helicopter experiment, the optimal control factor settings based on the highest SNR have been determined. These are shown in Table VII. In order to decide which level is better for maximising flight time, the SNR values at both low (level 1) and high (level 2) levels of each factor are compared. Once the optimal settings are established, it is useful to undertake a confirmation trial before onward actions are undertaken (Antony, 1996). Three helicopters were made using the optimal factor settings and the average flight time was recorded as 3.56 seconds. This shows an improvement of above 30 per cent on the average flight time using the range of variable settings. The results also reveal that flight time increases for larger wing length and smaller body length.

Summary and conclusions

The experiment was carried out with the aim of optimising the flight time of a paper helicopter. In order to study the effect of variables and the possible interactions between them in a minimum number of trials, the Taguchi approach to experimental design was adopted. As the experiment itself was simple, the students found it to be a clear illustration of the process of : defining the problem; identifying the control variables and possible interactions; defining the required levels for each variable/factor; determining the response of interest; selecting the most suitable orthogonal array; performing the experiment; undertaking the analysis; and interpreting the results to obtain a better understanding of the situation under review.
Table VII Optimal control factor settings Control factors Paper type Body length Wing length Body width Number of clips Wing shape Optimum level Regular (level 1) 8cm (level 1) 12cm (level 2) 2cm (level 1) 1 (level 1) Flat (level 1)

The Taguchi method is a powerful approach to address process variability and optimisation problems. However, the application of SDOE and Tm by the engineering fraternity in UK organisations is limited due, in part, to a shortage of skills in problem solving and inadequate statistical knowledge. This paper demonstrates a simple means of introducing students to this powerful tool. The approach uses a simple paper helicopter experiment. For simplicity, all control parameters were studied at two levels. This mirrors actual practice in most optimisation problems, factors at two levels are the most widely used (Gunst and Mason, 1991; Lucas, 1992). The paper helicopter experiment is quite old and has been widely used by many statisticians for teaching purposes. However, this approach has focused on minimal statistical jargon and number crunching and on the use of modern graphical tools to achieve a rapid understanding of the results from the statistical analysis. The authors strongly believe that the experiment provides a simple and beneficial way to help engineers approach experimental design in a way that ensures it is transferrable to their own work environment.

References
Antony, J. (1996), ``A strategic methodology to the use of advanced statistical quality control techniques, PhD thesis. Antony, J. (1998), ``Some key things industrial engineers should know about experimental design, Logistics Information Management, 1998, Vol. 11 No. 6, pp. 386-92. Antony, J. et al. (1996), ``Optimisation of core tube life using Taguchi experimental design methodology, Journal of Quality World (Technical Supplement), IQA, March, pp. 42-50. Antony, J. et al. (1998a), ``A strategic methodology to the use of advanced statistical quality improvement techniques, The TQM Magazine (The International Bi-Monthly for TQM), Vol. 10 No. 3, pp. 169-176. Antony, J. et al. (1998b), ``Key interactions, Journal of Manufacturing Engineer, IEE, Vol. 77 No. 3, pp. 136-8. Antony, J. et al. (1999), Experimental Quality A Strategic Approach to Achieve and Improve Quality, Kluwer Academic Publishers, Dordrecht, December. Bendell, A. (Ed.) (1989), Taguchi Methods Applications in World Industry, IFS Publications, Bedford. Daniel, C. (1959), ``Use of half-normal plots in interpreting factorial two level experiments, Technometrics, Vol. 1 No. 4, pp. 53-70.

148

Teaching the Taguchi method to industrial engineers Jiju Antony and Frenie Jiju Antony

Work Study Volume 50 . Number 4 . 2001 . 141149

Gunst, R.F. and Mason, R.L. (1991), How to Construct Fractional Factorial Experiments, ASQC Statistics Division, ASQC Press, Milwaukee, MI. Lucas, J.M. (1992), ``Split plotting and randomisation in industrial experiments, ASQC Quality Congress Transactions, Nashville, TN, pp. 374-82. Morrison, J.M. (1997), ``Statistical engineering the key to quality, Engineering Science and Education Journal, pp. 123-7. Phadke, M.S. (1989), Quality Engineering using Robust Design, Prentice-Hall International, Englewood Cliffs, NJ.

Ross, P.J., (1988), Taguchi Techniques for Quality Engineering, McGraw-Hill Publishers, New York, NY. Rowlands, H., Antony, J. and Knowles, G. (2000), ``An application of experimental design for process optimisation, The TQM Magazine, Vol. 12 No.2, pp. 78-83. Schmidt, S.R and Launsby, R.G. (1992), Understanding Industrial Designed Experiments, Air Academy Press, Washington, DC. Taguchi, G. (1986), Introduction to Quality Engineering, Asian Productivity Organisation, Tokyo.

Appendix
Table AI Coded design matrix of an L16 (21 5 ) orthogonal array Trial 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 1 1 1 1 2 2 2 2 1 1 1 1 2 2 2 2 3 1 1 1 1 2 2 2 2 2 2 2 2 1 1 1 1 4 1 1 2 2 1 1 2 2 1 1 2 2 1 1 2 2 5 1 1 2 2 1 1 2 2 2 2 1 1 2 2 1 1 6 1 1 2 2 2 2 1 1 1 1 2 2 2 2 1 1 7 1 1 2 2 2 2 1 1 2 2 1 1 1 1 2 2 Column 8 9 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 2 1 2 1 2 1 2 1 10 1 2 1 2 2 1 2 1 1 2 1 2 2 1 2 1 11 1 2 1 2 2 1 2 1 2 1 2 1 1 2 1 2 12 1 2 2 1 1 2 2 1 1 2 2 1 1 2 2 1 13 1 2 2 1 1 2 2 1 2 1 1 2 2 1 1 2 14 1 2 2 1 2 1 1 2 1 2 2 1 2 2 1 2 15 1 2 2 1 2 1 1 2 2 1 1 2 1 1 2 1

149