HANDLING UNCERTAIN INFORMATION IN WHOLE LIFE COSTING - A COMPARATIVE STUDY Mohammed Kishk, Assem Al-Hajj and Robert Pollock

Scott Sutherland School, The Robert Gordon University, Aberdeen AB10 7QB, UK. ABSTRACT A number of recently developed algorithms to handle uncertain information in whole life costing (WLC) are explained and validated in the context of two example applications. In the first example application, the proposed methodology is compared to the sensitivity analysis technique. The break-even point has been correctly identified in almost all cases. Furthermore, it has been shown how the employed fuzzy methodology may be seen as a generalised sensitivity approach to which has been added a measure of the precision with which input variables are known to the decision-maker. In the second example application, the proposed methodology is compared to two probabilistic techniques: the confidence index method and the Monte Carlo simulation (MCS) technique. The proposed methodology correctly identified the most uncertain cost items and portrayed well the confidence in ranking. Besides, all predicted net present values were in close agreement with those obtained by the MCS technique. This showed once more the robustness of various measures and concepts employed in the developed algorithms. 1.0 INTRODUCTION Whole life costing (WLC) is a technique that is used to facilitate the effective choice between a number of competing alternatives that differ not only in their initial costs but in their subsequent maintenance and operational costs as well. The method consists of two processes. First, all costs and revenues associated with the acquisition, use and maintenance and disposal of various alternatives are discounted to the present time. Then, these discounted costs of each alternative are summed up to calculate the net present value (NPV) of that alternative. In the second process, various alternatives are ranked in order of ascendant magnitude according to their NPVs and the best alternative is selected such that it has the lowest NPV. For results that are deterministic, there is no ambiguity in ranking the alternatives and the decision is straightforward. WLC, however, deals with the future and the future is unknown. Thus, it is crucial to risk assess the results before the final decision is made. In doing so, either a sensitivity analysis (SA) or a probabilistic risk assessment technique, usually the Monte Carlo simulation (MCS), is employed. The SA is used to identify the impact of a change in the value of a single risky independent parameter on the dependent variable, i.e. the NPV in a WLC exercise. The objective is usually to determine the break-even point defined as the value of the input-data element that causes the NPV of the least-cost alternative to equal that of the next-lowest-cost alternative (Kirk and DellIsola, 1995). The major advantage of the SA is that it explicitly shows the robustness of the ranking of alternatives (Flanagan and Norman, 1993, Woodward, 1995). However, it has two limitations. First, it is a univariate approach, i.e., only one parameter can be varied at a time. Thus, it is effective only when the uncertainty in one input-data element is predominant (Kirk and DellIsola, 1995). Secondly, it does not aim to quantify risk but rather to identify factors that are risk sensitive (Flanagan et al., 1989). Thus, it does not provide a definitive method of making the decision.

Kishk, Al-Hajj and Pollock

The MCS has been used in WLC modelling by many authors including Flanagan et al. (1987, 1989), Ko et al. (1998) and Goumas et al. (1999). In a MCS, every uncertain variable is represented by a probability distribution function (PDF). The resulting NPVs become random variables represented by PDFs. As noted by Flanagan et al. (1989), this provides the decision-maker with a wider view in the final choice between alternatives but will not remove the need for the decision-maker to apply judgement and there will be, inevitably, a degree of subjectivity in this judgement. Moreover, many researchers (Woodward, 1995; Chau, 1997; Byrne, 1997; Edwards and Bowen, 1998; among others) have criticized simulation techniques for their complexity and their expense in terms of computation time and expertise required to extract the knowledge. The confidence index (CI) is a simplified probabilistic approach. It is based on two assumptions (Kirk and DellIsola, 1995): (1) the uncertainties in all cost data are normally distributed; and (2) the high and low 90% estimates for each cost do in fact correspond to the true 90% points of the normal PDF for that cost. Obviously, the CI approach tackles some of the difficulties of MCS. However, the necessary assumption of normally distributed data and the above two restrictions limit its generality. The main assumption in probabilistic risk assessment techniques is that all uncertainties follow the characteristics of random uncertainty. This implies that all uncertainties are due to stochastic variability or to measurement or sampling error; and consequently are expressible by means of PDFs that are best derived from significant data. However, historic data for WLC within the construction industry is sparse (Bull, 1993; Ashworth, 1996, 1999; Wilkinson, 1996; Sterner, 2000, among others). Besides, facets of uncertainty in WLC data are not only random but also of a judgmental nature (Kishk, 2001). Due to the above limitations of existing risk assessment models in handling uncertain data in WLC modelling, the authors suggested that the fuzzy set theory (FST) might be more appropriate. This is mainly because it is easier to define fuzzy variables than random variables when no or limited information is available (Kaufmann and Gupta, 1985). Furthermore, mathematical concepts and operations within the framework of FST are much simpler than those within the probability theory (Ferrari and Savoia, 1998). In a series of papers (Kishk and Al-Hajj, 2000a, 2000b, 2000c, 2000d), the authors employed the FST to develop three models and algorithms to handle subjective assessments of input variables. In a subsequent paper (Kishk and Al-Hajj, 2001a), another algorithm has been developed to handle stochastic data and expert assessments as represented by probability density functions (PDFs) and fuzzy numbers (FNs), respectively, within the same model calculation. Recently, a WLC-based decision support algorithm has also been proposed (Kishk and Al-Hajj, 2001b). This algorithm systematically analyses uncertain input data and provides the decision-maker with a better impression of their validity and usability by the employment of two sets of measures. The first set includes two confidence measures CI1 and CI 2 , to evaluate the rank ordering of various competing alternatives. These factors may be interpreted as measures of the confidence in the two statements: A is better than B and A is at least as better as B, respectively, where A and B are two competing alternatives. The second set includes two uncertainty measures U and F to identify the significance of various costs regarding the ambiguity of the decision. In this paper, these algorithms are explained and compared to existing assessment techniques in the context of two example applications. This is done to highlight some of their unique features and to identify their relation to the SA, CI and MCS techniques.

Handling uncertain information in whole life costing

2.0 EXAMPLE (1) In this section, an example problem given in Kirk and DellIsola (1995) is solved. A student recreation centre including a gymnasium was to be designed and it was required to undertake a feasibility study for the use of a solar-assisted domestic hot water system. Cost data of the two design alternatives are summarised in Table (1). Because the uncertainty of the economic life of the solar-energy system, T , a SA of varying its value from 10 to 25 with a best estimate of 18 years using a discount rate of 7% has been requested. Kirk and DellIsola (1995) solved the problem by calculating the present worth savings from using the solar energy system and a break-even point of 16 years was obtained. To ease the comparison with the proposed algorithms, the problem is resolved by calculating the net present values of both alternatives. Again, the break-even point is 16 years and the corresponding net present value is $144,723 as shown in Fig. (1). Table (1): Cost data for example (1). Cost Initial construction cost Annual Maintenance and repair costs Annual fuel cost Salvage value
180

Fuel System Baseline Baseline $ 15,320 0

Solar System $ 129,000 $ 2,720 0 $ 29,440

Net Present Value ($1000s)

160

144.723
140

120

Solar. Fuel.
100 8 12 16 20 24 28

Figure (1): The sensitivity test using the net present values. The net present value (NPV) algorithm (Kishk and Al-Hajj, 2000b) was also employed to solve this example problem. Low, best and high estimates were used to define the membership function of the life cycle, T . Five different MFs were considered: a triangular fuzzy number (TFN), a trapezoidal fuzzy number (TrFN) and three normal PDFs. Figure (2) depicts the resulting NPVs of both alternatives for the TFN and TrFN cases. For the TFN case, the same break-even point is calculated by the algorithm at net present value of $144,723 with a membership value of 0.75. This value corresponds to an 3

Kishk, Al-Hajj and Pollock

economic life of 16 years (Fig. 3) as previously calculated by the sensitivity approach. For the TrFN case, the resulting MFs of the net present values are two crisp sets (intervals) whose boundaries are defined by the NPVs of both alternatives at the corresponding boundaries of the TrFN, i.e. [10, 25] years. Thus, a break-even interval, [$133,138, $155,274] , was obtained. The true break-even point of $144,723 is included in this interval as shown in Fig. (2). Obviously, a single break-even point could not be obtained because the use of a TrFN implies that the low, best and high estimates were given the same membership value, i.e. = 1 .

1.0
Solar Fuel TFN TrFN

0.8

0.750

0.6

0.4 0.2 0.0 100 120 140 144.723 160 180

NPV ($1000s) Figure (2): The net present values for the TFN and TrFN cases (example 1).

1.0

0.8

0.750

0.6

0.4 0.2 0.0 8 12 16 20 24 28

T, years Figure (3): Break-even point for the TFN case (example 1).

Handling uncertain information in whole life costing

To further investigate the effect of the shape of the membership function, three normal PDFs have been used to model the economic life of the solar system with the best estimate of 18 years as the mean value and standard deviations, , of 1, 1.5 and 2. Figure (4) shows the resulting NPVs for these three cases. In all cases, the correct break-even point of $144,723 was predicted with associated membership values of 0.411, 0.674 and 0.804 for = 1, 1.5 and 2 , respectively. Again, these membership values correspond to the same economic life of 16 years as shown in Figure (5). This shows the robustness of the transformation algorithm.

1.0

0.8

0.804 0.674

Solar Fuel PDF, = 1.0 PDF, = 1.5 PDF, = 2.0

0.6

0.4

0.411

0.2

0.0 100 120 140

144.723

160

180

NPV ($1000s) Figure (4): MFs for the net present values for PDF cases (example 1).

1.0

0.8

0.804 0.674

= 1.0 = 1.5 = 2.0

0.6

0.4

0.411

0.2

0.0 8 12 16 20 24 28

T, years Figure (5): Break-even points for PDF cases (example 1).

Kishk, Al-Hajj and Pollock

Figure (4) shows also how the uncertainty in information is reflected in the predicted solution as the spread of the calculated NPVs increases with the uncertainty of data, i.e. as the value of the standard deviation increases. As expected, the spread is zero for the certain case, = 0 , shown with solid lines in the figure. In the TrFN case, the fuel system was ranked first, while the solar-assisted system was recommended in all other cases. The confidence measure in ranking for all studied cases are summarised in Table (2). All these measures are in general agreement with common sense which shows the effectiveness of the ranking algorithm and the employed confidence measures.
Table (2): Confidence measures of ranking (example 1). Case

Crisp Normal PDF, = 1.0 Normal PDF, = 1.5 Normal PDF, = 2.0 TFN TrFN

CI 1 1.000 0.550 0.434 0.369 0.342 0.274

CI 2 1.000 0.743 0.656 0.600 0.572 0.512

3.0 EXAMPLE (2) Cost estimates of two alternative floor finishes for an administrative facility are summarised in Table (3). It is required to choose the best option for an analysis period of 25 years and a discount rate of 10%. Kirk and DellIsola (1995) used the confidence index approach to solve this example. Alternative L was ranked first with a confidence index of 0.1498 which indicates low confidence in the choice of alternative L for implementation. Besides, their computations indicated also that the high uncertainty of annual M&R are the main cause to this low confidence in ranking. Because the CI approach does not give detailed NPVs, this problem is re-solved using MCS using the Crystal Ball 2000 simulation software (Decisioneering, 2000) as shown in Fig. (6). The fuzzy NPV algorithm was also used and its results are depicted with the thick curves in Fig. (7). Results of the MCS and fuzzy solutions are summarised in Table (4). To allow for a clearer comparison between the MCS and fuzzy results, the relative frequencies in Fig. (6) were transformed to possibility distributions (Kishk and Al-Hajj, 2001a) and are plotted in Fig. (7) with thin curves. As shown, fuzzy results agreed well with those obtained from MCS. Table (3): Cost estimates for example (2). Alternatives 1. Alternative (L): Initial construction costs Maintenance and repair annual cost. 2. Alternative (H): Initial construction costs Maintenance and repair annual cost. High Estimates Best Low

$ 7,600 $ 2,700 $ 14,500 $ 2,000

$ 7,200 $ 1,800 $ 14,500 $ 1,200

$ 6,800 $ 700 $ 13,700 $ 400

Alternative L was ranked first and the confidence measures in this ranking are summarised in Table (5). These relatively low measures reflect low confidence in this ranking. Besides, annual M&R costs were identified to have the dominant contribution to the ambiguity of ranking as clearly indicated by the calculated measures of uncertainty in Table (6). These results are in general agreement with those obtained by the confidence index approach.

Handling uncertain information in whole life costing

0.04
Alternative L. Alternative H.

8000

0.02

4000

0.01

2000

0.00 12 16 20 24 28 32 36

NPV ($1000s)

Figure (6): Simulation results (example 2).

1.0

0.8

0.6

0.4
Alternative L. Alternative H.

0.2

0.0 12 16 20 24 28 32 36

NPV ($1000s)

Figure (7): MFs of Net present values (example 2 ) Table (4): Summary of results (example 2).

Alternative L H

MCS Minimum $14,992 $17,345 Maximum $32,095 $33,407 Mean $23,535 $25,407 7 Minimum $14,967 $17,331

Fuzzy Maximum $32,105 $33,454 Removal $23,536 $25,392

Frequency (200,000 trials)

0.03

6000

Relative frequency

Kishk, Al-Hajj and Pollock

Table (5): Measures of confidence (example 2).

Rank 1 2

Alternatives Alternative L Alternative H

Alternative L CI2 CI1 ----0.000 0.434

Alternative H CI1 CI2 0.132 0.566 -----

Table (6): Measures of uncertainty (example 2).

Cost and value items (discounted & normalised) Initial cost Maintenance and repair annual costs

Alternative H U F 0.000 0.000 0.560 0.307

Alternative L U F 0.007 0.018 0.598 0.335

Kirk and DellIsola re-solved this example problem using closer estimates for annual M&R costs (Table 7). Alternative L was ranked first with a confidence index of 0.259 indicating high confidence in the choice of alternative L for implementation. Again, both the Crystal Ball 2000 simulation software (Decisioneering, 2000) and the NPV algorithms were used to obtain the revised NPVs of both alternatives. These revised results are summarised in Table (8). Figure (8) shows the results of the simulation exercise (200,000 trials) and their equivalent MFs are depicted with the thin lines in Fig. (9). As shown in Fig. (9) and Table (8), the revised results obtained from the proposed fuzzy algorithms agreed well with those obtained from MCS and are slightly more conservative. Again, alternative L was ranked first and the confidence measures in this ranking are summarised in Table (9). These relatively high measures reflect higher confidence in this ranking as previously indicated by the confidence index technique. These results indicate again the robustness of various measures employed in the proposed algorithms.
Table (7): Revised cost estimates (example 2). Alternatives 1. Alternative (L): Initial construction costs Maintenance and repair annual cost. 2. Alternative (H): Initial construction costs Maintenance and repair annual cost. High Estimates Best Low

7,600 2,300 14,500 1,800

7,200 1,800 14,500 1,200

6,800 1,300 13,700 600

Table (8): Summary of revised results (example 2).

Alternative L H

MCS Minimum Maximum $18,612 $28,464 $19,151 $31,630

Mean $23,533 $25,394

Fuzzy Minimum Maximum $18,597 $28,474 $19,146 $31,639

Removal $23,536 $25,392

Table (9): Revised measures of confidence (example 2).

Rank 1 2

Alternatives Alternative L Alternative H

Alternative L CI2 CI1 ----0.0003 0.402

Alternative H CI1 CI2 0.196 0.598 -----

Handling uncertain information in whole life costing

0.04
Alternative L. Alternative H.

8000

0.02

4000

0.01

2000

0.00 16 20 24 28 32

NPV ($1000s)

Figure (8): Simulation results (example 2, revised case).

1.0

0.8

0.6

0.4
Alternative L. Alternative H.

0.2

0.0 16 20 24 28 32

NPV ($1000s)

Figure (9): MFs of NPVs (example 2, revised case). 4.0 CONCLUSIONS AND FUTURE RESEARCH

A number of recently developed algorithms have been illustrated using two case studies. Typical results showed a well agreement with traditional risk assessment techniques. Besides, smooth output MFs have been obtained despite the relatively small number of intervals employed, illustrating the robustness and computational efficiency of the implemented algorithms. Furthermore, these algorithms not only have the desirable features of existing techniques but also have more advantages as follows. 9

Frequency (200,000 trials)

0.03

6000

Relative frequency

Kishk, Al-Hajj and Pollock

Compared to the SA technique, they have the desirable property of transparency. Besides, they can deal with the uncertainty of multiple variables simultaneously rather than being a univariate approach. Compared to the MCS technique, they have the desirable property of producing the distribution of all possible values of the output variable but in a less complex and a more effective way. Compared to the CI approach, they have the desirable property of ranking competing alternatives and giving confidence measures in this ranking but they are not limited to a specific type of distributions of input variables.

Because of these unique features, these algorithms will be employed to develop IT applications of whole life costing to support life-cycle decision-making in the design and management of construction assets. Part of this objective will be achieved through an ongoing EPSRC funded research project (Al-Hajj et al., 2001; Aouad et al., 2001) undertaken by a joint collaboration with another team from the University of Salford.
5.0 REFERENCES Al-Hajj, A., Pollock, R., Kishk, M., Aouad, G., Sun, M. and Bakis, N. (2001) On the requirements for effective information management in whole life costing within an integrated environment. Proceedings of the Annual Conference of the RICS Research Foundation (COBRA2001), Glasgow Caledonian University, 3-5 September, 2, 402-413. Aouad, G., Bakis, N., Amaratunga, D., Osbaldiston, S., Sun, M., Kishk, M., Al-Hajj, A. and Pollock, R. (2001) An integrated life cycle costing database a conceptual framework. Proceedings of the 17th Annual Conference of the Association of Researchers in Construction Management (ARCOM2001), University of Salford, 5-7 September, 1, 421-431. Ashworth, A. (1996) Estimating the life expectancies of building components in life cycle costing calculations, Structural Survey, 14 (2), 4-8. Ashworth, A. (1999) Cost studies of buildings, Longman. Bull, J. W. (1993) The way ahead for life cycle costing in the construction industry. In Bull, J. W. (ed.) Life Cycle Costing for Construction. Blackie Academic & Professional, Glasgow, UK. Byrne, P. (1997) Fuzzy DCF: a contradiction in terms, or a way to better investment appraisal? Proceedings of Cutting Edge 97, RICS. Chau, K. W. (1997) Monte Carlo simulation of construction costs using subjective data: response, Construction Management and Economics, 15, 109-115. Decisioneering (2000) Crystall Ball 2000 User Manual. Decisioneering, Denver, USA. Edwards, P.J. and Bowen, P. A. (1998) Practices, barriers and benefits of risk management process in building services cost estimation: comment, Construction Management and Economics, 16, 105-108. Goumas, M. G., Lygerou, V. A. and Papayannakis, L. E. (1999) Computational methods for planning and evaluating geothermal energy projects, Energy Policy, 27, 147154. Flanagan, R., Kendell, A., Norman, G. and Robinson, G. (1987) Life Cycle Costing and Risk Management, Construction Management and Economics, 5, 53-71. Flanagan, R., Norman, G., Meadows, J. and Robinson, G. (1989) Life Cycle Costing Theory and Practice. BSP Professional Books.

Handling uncertain information in whole life costing

Flanagan, R. and Norman, G. (1993) Risk management and construction. Blackwell Scientific Publications. Kirk, S., J. and DellIsola, A. J. (1995) Life Cycle Costing for Design Professionals. McGrew-Hill Book Company, New York. Kishk, M. (2001) An Integrated fuzzy approach to whole life costing based decision making. Unublished PhD Thesis. Scott Sutherland School, The Robert Gordon University, Aberdeen. Kishk M. and Al-Hajj A. (2000a) A fuzzy approach to model subjectivity in life cycle costing. Proceedings of the BF2000 national conference of postgraduate research in the built and human environment, The University of Salford, 9-10 March, 270280. Kishk M. and Al-Hajj A. (2000b) A fuzzy model and algorithm to handle subjectivity in life cycle costing based decision-making. Journal of financial management of property and construction, 5, 1-2, August, 93-104. Kishk M. and Al-Hajj A. (2000c) Handling linguistic assessments in life cycle costing - a fuzzy approach. Proceedings of the construction and building research conference of the RICS research foundation (COBRA2000), The University of Greenwich, 30 August - 1 September, 2000, 228-243. Kishk M. and Al-Hajj A. (2000d) Modelling of life cycle costs of alternatives with different lives. Proceedings of the 16th annual conference of the Association of Researchers in Construction Management (ARCOM2000), Glasgow Caledonian University, 6-8 September, 577-586. Kishk M. and Al-Hajj A. (2001a) Integrating Subjective and Stochastic Data in Life Cycle Costing Calculations. Proceedings of the first international postgraduate research conference in the Built and Human Environment, University of Salford, 15-16 March, 329-344. Kishk, M.; and Al-Hajj, A. (2001b) An innovative approach to integrating the analysis of uncertainty into life cycle costing. Proceedings of the first international conference of Innovation in Architecture, Engineering and Construction, Loughborough University, 19-20 July, 10-21. Ko, W. I., Choi, J. W., Kang, C. H. and Lee, J. S. (1998) Nuclear fuel cycle cost analysis using a probabilistic simulation technique, Annals of Nuclear Energy, 25, 771789. Sterner, E. (2000) Life-cycle costing and its use in the Swedish building sector, Building Research & Information, 28 (5/6), 387-393. Wilkinson, S. (1996) Barriers to LCC Use in the New Zealand Construction Industry. Proceedings of the 7th International Symposium on Economic Management of Innovation, Productivity and Quality in Construction, Zagreb, 447-456. Woodward, D. G. (1995) Use of sensitivity analysis in Build-Own-Operate-Transfer project evaluation, International Journal of Project Management, 13 (4), 239-246.

11