CROSS REFERENCE TO RELATED APPLICATIONS

[0001]
This application claims priority from the provisional application Serial No. 60/377,350, filed Apr. 30, 2002, entitled “Methods and Apparatus for Decision Making”.
FIELD OF THE INVENTION

[0002]
The present invention relates to methods and apparatus for decision making, including software therefor.
BACKGROUND OF THE INVENTION

[0003]
Decision making is a most common human activity. Individuals and organisations make all kinds of decisions in a variety of ways on a regular basis. Most decision problems are associated with a number of criteria, which may be measured in different ways, be in conflict with one another, and comprise both a quantitative and qualitative nature. In many situations, decision makers may have to make decisions on the basis of incomplete or partial information. For instance, buying a car may be an individual or a family decision and a customer will not buy a car without taking into account several criteria such as price, safety measures, size of engine, and general quality. Similarly, a company often will not do business with a supplier without assessing many criteria such as financial stability, technical capability, quality and after sales services.

[0004]
There is a large literature associated with decision sciences, in which techniques for aiding or actually making decisions are proposed. Of most relevance to the present application is Multiple Criteria Decision Analysis (MCDA), which is an important area of decision sciences wherein scientific methods are investigated and developed in order to support decision making with multiple criteria.

[0005]
A decision associated with multiple criteria is deemed to be properly made if all criteria in conflict are properly balanced and sufficiently satisfied. A MCDA problem can be generally modelled using a decision matrix, where a column represents a criterion, a row an alternative decision, and an element the outcome of a decision on a criterion. The decision matrix for a car selection problem, for example, may look like Table 1.
TABLE 1 


Example of Decision Matrix 
     General 
 Engine Size  Fuel Consumption  Price  . . .  Quality 
 
Car 1  1400 cc  40 miles/gallon  £8,000  . . .  Good 
Car 2  1500 cc  45 miles/gallon  £9,000  . . .  Excellent 
. . .  . . .  . . .  . . .  . . .  . . . 
Car N  1100  47 miles/gallon  £7,000  . . .  Good 


[0006]
Several methods have been proposed to deal with MCDA problems represented in the form of a decision matrix. Multiple criteria utility (value) function (MCUF) methods are among the simplest and most commonly used (see, for example, E. JacquetLagreze and J. Siskox, “Assessing a set of additive utility functions for multicriteria decision making, the UTA method”, European Journal of Operational Research, Vol. 10, pp. 151164, 1982, and R. L. Keeney and H. Raiffa, Decision with Multiple Objectives: Preference and Value Tradeoffs, John Wiley and Sons, New York, 1976).

[0007]
The MCUF methods are based on the estimation of utility for each outcome in a decision matrix. However, if a MCDA problem involves a large number of criteria and alternative decisions, estimating the utilities of all outcomes at every alternative on each criterion will become a tedious procedure and as such the MCUF methods will be difficult to apply in a satisfactory way (T. J. Stewart, “A critical survey on the status of multiple criteria decision making theory and practice”, OMEGA International Journal of Management Science, Vol. 20, No. 56, pp. 569586, 1992).

[0008]
Pairwise comparisons between pairs of criteria were primarily used to estimate relative weights of criteria in several methods including the eigenvector method (T. L. Saaty, The Analytic Hierarchy Process, University of Pittsburgh, 1988), the geometric least square method (G. Islei and A. G. Lockett, “Judgmental modelling based on geometric least squares”, European Journal of Operational Research, Vol.36, No. 1, pp.2735, 1988) and the geometric mean method. Pairwise comparison matrices have also been used to assess alterative decisions with respect to a particular criterion such as in Analytical Hierarchy Process (AHP) (Saaty, ibid) and in judgmental modelling based on the geometric least square method (Islei and Locket, ibid). However, using pairwise comparisons to assess alternatives may lead to problems such as rank reversal as within the AHP framework (V Belton and T Gears “On a shortcoming of Saaty's method of analytic hierarchy”, OMEGA, vol. 11, No. 3, pp 228230, 1981; Stewart, ibid). These difficulties have lead to a long debate on how quantitative and qualitative assessments should be modelled and aggregated. Furthermore, both MCUF and AHP methods are incapable of properly coping with decision problems with missing information. If assessment information is missing for one criterion, one has to either abandon this criterion altogether or make assumptions, ie to use fabricated information. However, this may mislead the decision making process.

[0009]
Fuzzy sets based methods have been developed to deal with MCDA problems with uncertainties. The main feature of such methods is their capability of handling subjective judgements in a natural manner. Therefore, they provide attractive frameworks to represent qualitative criteria and model human judgements (R R Yager “Decisionmaking under various types of uncertainties”, Journal of Intelligent and Fuzzy Systems, Vol.3, No. 4, pp 317323, 1995). However, fuzzy set methods suffer from two fundamental drawbacks. Firstly, they use a simplistic approach and limited linguistic variables to model a variety of information including both precise numbers and imprecise judgements. The consequences of this modelling strategy include the loss of precision in describing precise data and the lack of flexibility in capturing the diversity of information. The second drawback results from the use of fuzzy operations for criteria aggregation. Traditional fuzzy operators may lead to the loss of information in the process of aggregating a large number of criteria (J Wang, J B Yang and P Sen “Safety analysis and synthesis using fuzzy sets and evidential reasoning”, Reliability Engineering and Systems Safety, Vol. 47, No. 2, pp 103118, 1995).

[0010]
The present inventors have developed a MCDA method which has been termed evidential reasoning (ER) (see J. Wang, J. B. Yang and P. Sen, “Safety analysis and synthesis using fuzzy sets and evidential reasoning”, Reliability Engineering and System Safety, Vol. 47, No. 2, pp. 103118, 1995, J. B. Yang and M. G. Singh, “An evidential reasoning approach for multiple attribute decision making with uncertainty”, IEEE Transactions on Systems, Man and Cybernetics, Vol. 24, No. 1, pp. 118, 1994; J. B. Yang and P. Sen, “A general multilevel evaluation process for hybrid MADM with uncertainty”, IEEE Transactions on Systems, Man, and Cybernetics, Vol. 24, No. 10, pp. 14581473, 1994; and Z. J. Zhang, J. B. Yang and D. L. Xu, “A hierarchical analysis model for multiobjective decision making”, in Analysis, Design and Evaluation of ManMachine System 1989, Selected Papers from the 4th IFAC/IFIP/IFORS/IEA Conference, Xian, P. R. China, September 1989, Pergamon, Oxford, UK, pp.1318, 1990).

[0011]
In the ER approach, it is proposed to use the concept of belief degrees in an assessment framework to model subjective judgements and develop an evidential reasoning algorithm to aggregate criteria in the assessment framework (Zhang, Yang and Xu; ibid: Yang and Singh, ibid; Yang and Sen, ibid). Compared with fuzzy sets methods, the ER approach provides a more flexible way of modelling human judgements (Yang and Sen, ibid) and the ER criteria aggregation process is also based on the rigorous DempsterShafer theory of evidence (G. A. Shafer, Mathematical Theory of Evidence, Princeton University Press, Princeton, USA, 1976, the contents of which, together with the contents of the other publications cited above, are hereby incorporated by reference). However, the prior art ER technique as described in the above mentioned publications is primarily of academic interest, since it is unable to properly accommodate a variety of “real life” situations. For example, the prior art technique is not capable of accommodating precise data or properly handling incomplete information, which may be caused due to a lack of information, the complexity of a decision problem and the inability of humans to provide precise judgements. Also, the old ER algorithm does not provide a rigorous process of aggregating incomplete information.

[0012]
Therefore, there is a need to provide an improved MCDA technique which is capable of dealing with “real life” situations, and of overcoming the above described problems associated with the prior art.
SUMMARY OF THE INVENTION

[0013]
The present invention addresses the aforesaid need, and overcomes the above described problems. The present invention provides a rigorous means to support in a practical way the solution of MCDA problems. It is capable of dealing with quantitative and qualitative information, and can handle imprecise subjective information in a way that is consistent and reliable.

[0014]
According to a first aspect of the invention there is provided a multiple criteria decision analysis method in which a plurality L of basic criteria are assessed in order to assess a general criterion, comprising the steps of:

[0015]
making an assessment {(K_{m,l},γ_{m.l}),m=1, . . . ,M} of the l^{th }basic criteria under a set of grades {K_{m,l}, m=1, . . . , ml;

[0016]
and transforming the assessment {K
_{m,l},γ
_{m,l},m=1, . . . ,M} to an assessment {(H
_{n},β
_{n,l}),=n=1, . . . , N} of the general criterion under a set of grades {H
_{n},n=1, . . . ,N} using the matrix equation:
$\left[\begin{array}{c}{\beta}_{1,l}\\ {\beta}_{2,l}\\ \vdots \\ {\beta}_{N,l}\end{array}\right]=\left[\begin{array}{cccc}{\alpha}_{1,1}& {\alpha}_{1,2}& \dots & {\alpha}_{1,M}\\ {\alpha}_{2,1}& {\alpha}_{2,2}& \dots & {\alpha}_{2,M}\\ \vdots & \vdots & \u22f0& \vdots \\ {\alpha}_{N,1}& {\alpha}_{N,2}& \dots & {\alpha}_{N,M}\end{array}\right]=\left[\begin{array}{c}{\gamma}_{1,l}\\ {\gamma}_{1,l}\\ \vdots \\ {\gamma}_{M,l}\end{array}\right]$

[0017]
wherein:

[0018]
H_{n }is the n^{th }grade for assessment of the general criterion;

[0019]
K_{m,l }is the m^{th }grade for assessment of the l^{th }basic criterion;

[0020]
α_{n,m }is the degree to which K_{m,l }implies H_{n};

[0021]
γ_{m,l }is the degree to which the l^{th }basic criterion is assessed to K_{m,l};

[0022]
β
_{n,l }is the degree to which the l
^{th }basic criterion is assessed to Hn; and
$\sum _{n=1}^{N}\ue89e{\alpha}_{n,m}=1$

[0023]
This method provides a rule based transformation of qualitative information. The values of α_{n,m}(m=1, . . . ,M and n=1, . . . ,N) maybe assigned by a decision maker. The decision maker may be, for instance, the user of software embodying the method. A grade K_{m,l }may imply a grade H_{n }to a degree of α_{n,m }wherein m=1, . . . ,M.

[0024]
The values of α
_{n,m }(m=1, . . . , M and n=1, . . . ,N) may be determined using the following equations:
${\alpha}_{n,m}=\frac{u\ue8a0\left({H}_{n+1}\right)u\ue8a0\left({K}_{m,l}\right)}{u\ue8a0\left({H}_{n+1}\right)u\ue8a0\left({H}_{n}\right)},{\alpha}_{n+1,m}=1{\alpha}_{n,m},$

[0025]
α_{i,m}=0(i=1, . . . ,N,i≠n,n+1)

[0026]
if u(H_{n})≦u(K_{m,l})≦u(H_{n+1}) for n=I, . . . ,N−1;m=1, . . . ,M

[0027]
where u(H_{n}) and u(K_{m,l}) are utilities of H_{n }and K_{m,l}, respectively. This can be regarded as a utility based information transformation. It can be used, for example, if the user of the method does not want or is unable to enter values for α_{n,m}.

[0028]
At least one of u(H_{n}) and u(K_{m,l}) may be estimated by a decision maker.

[0029]
At least one of u(H
_{n}) and u(K
_{m,l}) may be determined using one or more of the equations:
$u\ue8a0\left({H}_{n}\right)=\frac{n1}{N1}\ue89e\text{\hspace{1em}}\ue89e\mathrm{for}\ue89e\text{\hspace{1em}}\ue89en=1,\dots \ue89e\text{\hspace{1em}},N\ue89e\text{\hspace{1em}}\ue89e\mathrm{if}\ue89e\text{\hspace{1em}}\ue89e{H}_{n+1}$

[0030]
is preferred to H
_{n};
$u\ue8a0\left({K}_{m,l}\right)=\frac{m1}{M1}\ue89e\text{\hspace{1em}}\ue89e\mathrm{for}\ue89e\text{\hspace{1em}}\ue89em=1,\dots \ue89e\text{\hspace{1em}},M\ue89e\text{\hspace{1em}}\ue89e\mathrm{if}\ue89e\text{\hspace{1em}}\ue89e{K}_{m+1,l}$

[0031]
is preferred to K_{m,l}.

[0032]
According further to the method:

[0033]
a basic criterion may be assessed using the set {(k_{j},p_{j}), j=1, . . . ,P}

[0034]
where k
_{j }is a number and p
_{j }is the probability of the basic criterion taking the number k
_{j }and the values of γ
_{m,l }in the assessment {(k
_{m,l},γ
_{m,l}),m=1, . . . ,M} are calculated using the equation
${\gamma}_{m,l}=\sum _{j=1}^{P}\ue89e{s}_{m,j}\ue89e{p}_{j}\ue89e\text{\hspace{1em}}\ue89e\mathrm{for}\ue89e\text{\hspace{1em}}\ue89em=1,\dots \ue89e\text{\hspace{1em}},M$

[0035]
wherein:
${s}_{m,j}=\frac{{K}_{m+1,l}{k}_{j}}{{K}_{m+1,l}{K}_{m,l}},{s}_{m+1,j}=1{s}_{m,j},{s}_{i,j}=0;$

(m=1, . . . ,M−1); j=1, . . . , P; i=1, . . . , M, i≠m, m+1 if K _{m,l} ≦k _{j} ≦K _{m+1,l }

[0036]
In this way, γ_{m,l }can be calculated for quantitative criteria, and a quantitative data transformation can be performed. γ_{m,l }may be assigned directly by a decision maker in the case of qualitative criteria.

[0037]
The first aspect of the invention may be performed in conjunction with the second aspect of the invention, ie, a MCDA method may comprise both the first and second aspects of the invention.

[0038]
According to a second aspect of the invention there is provided a multiple criteria decision analysis method in which a plurality L of basic criteria are assessed in order to assess a general criterion, comprising the steps of:

[0039]
assigning weights W_{i}(i=1, . . . L) to the L basic criteria;

[0040]
normalising the weights using the equations
${\omega}_{i}=\frac{{W}_{i}}{\sum _{j=1}^{L}\ue89e{W}_{j}}\ue89e\left(i=1,\dots \ue89e\text{\hspace{1em}}\ue89eL\right);$

[0041]
determining β_{n,i}, wherein β_{n,i }is the degree to which the i^{th }basic criterion is assessed to H_{n}, and H_{n }is the nth grade for assessment of the general criterion, the general criterion being assessed into N grades;

[0042]
calculating weighted degrees of belief m_{n,i }from the equation

m _{n,i} =ω _{i}β_{n,i},(n=1, . . . , N;i=1, . . . ,L);

[0043]
and

[0044]
calculating a remaining probability mass m
_{H,i }from the equation
${m}_{H,i}=1\sum _{n=1}^{N}\ue89e{m}_{n,i}\ue8a0\left(i=1,\dots \ue89e\text{\hspace{1em}},L\right).$

[0045]
In this way, it is possible to deal with incomplete information. The weights may be assigned directly by a decision maker or estimated.

[0046]
m_{H,i }may be decomposed into {overscore (m)}_{H,i }and {tilde over (m)}_{H,i}, wherein:

[0047]
m_{H,i}={overscore (m)}_{H,i}+{tilde over (m)}_{H,i};

[0048]
{overscore (m)}
_{H,i}=1−ω
_{i}; and
${\stackrel{~}{m}}_{H,i}={\omega}_{i}\ue8a0\left(1\sum _{n=1}^{N}\ue89e{\beta}_{n,i}\right)\ue89e\text{\hspace{1em}}\ue89e\mathrm{for}\ue89e\text{\hspace{1em}}\ue89ei=1,\dots \ue89e\text{\hspace{1em}},L.$

[0049]
This approach permits greatly advantageous treatment of incomplete information.

[0050]
m_{n,i}, {overscore (m)}_{H,i}, and {tilde over (m)}_{H,i}(i=1, . . . ,L) maybe aggregated into combined probability masses I_{n,L}, {overscore (I)}_{H,L }and Ĩ_{H,L }respectively, using equations i) to ix) in a recursive manner

I_{n,1=m} _{n,l}(n=1,2, . . . ,N) i)

I_{H,l}=m_{H,l} ii)

Ĩ_{H,l}={tilde over (m)}_{H,l} iii)

{overscore (I)}
_{H,l}={overscore (m)}
_{H,l} iv)
$\begin{array}{cc}{K}_{i+1}={\left[1\sum _{t=1}^{N}\ue89e\sum _{j\ue89e\underset{j\ne t}{=}\ue89e1}^{N}\ue89e{I}_{t,i}\ue89e{m}_{j,i+1}\right]}^{1}& v)\end{array}$
I
_{n,i+1} =K _{i+1} [I _{n,i} m _{n,i+1} +I _{H,i} m _{n,i+1} +I _{n,i} m _{H,i+1}](
n= 1,2 . . . , N) vi)

Ĩ_{H,i+1} =+K _{i+1} [Ĩ _{H,i} {tilde over (m)} _{H,i+1} +{overscore (I)} _{H,i+1} {tilde over (m)} _{H,i+1} +Ĩ _{H,i+1} {overscore (m)} _{H,i+1}] vii)

{overscore (I)}_{H,i+1} =K _{i+1} [{overscore (H)} _{H,i+1} {overscore (m)} _{H,i+1}] Viii)

I _{H,i+1} ={overscore (I)} _{H,i+1} +{overscore (I)} _{H,i+1} ix)

[0051]
i={1,2, . . . ,L−1}

[0052]
This permits the probability masses to be combined, and allows upper and lower bounds of probability masses to be provided. From this, ranges of combined assessments can be generated.

[0053]
Combined degrees of belief β
_{n}, and γ
_{H }may be generated using the equations:
${\beta}_{n}=\frac{{I}_{n,L}}{1{\stackrel{\_}{I}}_{H,L}}\ue89e\text{\hspace{1em}}\ue89en=1,2,\dots \ue89e\text{\hspace{1em}},N$ ${\beta}_{H}=\frac{{\stackrel{~}{I}}_{H,L}}{1{\stackrel{\_}{I}}_{H,L}}$

[0054]
wherein β_{n }is a degree of belief to which the general criterion is assessed to the n^{th }grade H_{n }and β_{H }is a remaining degree of belief which is not assigned to any specific grade.

[0055]
Further according to the method wherein each grade H
_{n+1 }is more favourable than H
_{n}, performance indicators of a general criterion may be generated using the equations:
${u}_{\mathrm{max}}=\underset{n=1}{\sum ^{N1}}\ue89e{\beta}_{n}\ue89eu\ue8a0\left({H}_{n}\right)+\left({\beta}_{N}+{\beta}_{H}\right)\ue89eu\ue8a0\left({H}_{N}\right)$ ${u}_{\mathrm{min}}=\left({\beta}_{1}+{\beta}_{H}\right)\ue89eu\ue8a0\left({H}_{1}\right)+\sum _{n=2}^{N}\ue89e{\beta}_{n}\ue89eu\ue8a0\left({H}_{n}\right).\text{}\ue89e{u}_{\mathrm{avg}}=\frac{{u}_{\mathrm{max}}+{u}_{\mathrm{min}}}{2}$

[0056]
wherein u
_{max}, u
_{min }and u
_{avg }are the best possible, worst possible and average performance indicators respectively, and u(H
_{n}) (n=1, . . . N) is optionally defined by
$u\ue8a0\left({H}_{n}\right)=\frac{n1}{N1}.$

[0057]
Alternatively, u(H_{n}) may be estimated by a decision maker.

[0058]
The values of β_{n,i }may be determined using the method of the first aspect of the invention.

[0059]
According to a third aspect of the invention there is provided a carrier medium storing a computer program, which computer program performs the method of the first aspect of the invention. In this instance, a decision maker may be a user of the computer program.

[0060]
According to a fourth aspect of the invention there is provided a carrier medium storing a computer program, which computer program performs the method of the second aspect of the invention.

[0061]
According to a fifth aspect of the invention there is provided a computer adapted to perform the method of the first aspect of the invention.

[0062]
According to a sixth aspect of the invention there is provided a computer adapted to perform the method of the second aspect of the invention.

[0063]
Methods, computer programs and carrier media therefor in accordance with the invention will now be described with reference to the accompanying drawings.
BRIEF DESCRIPTION OF THE DRAWINGS

[0064]
[0064]FIG. 1 is a schematic diagram of quality criteria for a motor engine;

[0065]
[0065]FIG. 2 is a schematic diagram of general and basic criteria;

[0066]
[0066]FIG. 3 shows the main window produced by an exemplary computer program;

[0067]
[0067]FIG. 4 shows a software driven interface for implementing a rule based quantitative data transformation technique;

[0068]
[0068]FIG. 5 shows a software driven interface for implementing a rule based qualitative information transformation technique;

[0069]
[0069]FIG. 6 shows a software driven interface which supports utility estimation;

[0070]
[0070]FIG. 7 shows a software driven interface which enables random numerical data to be inputted;

[0071]
[0071]FIG. 8 shows a software driven interface which permits a user to assign degrees of belief;

[0072]
[0072]FIG. 9 shows a graphical display of a distributed assessment;

[0073]
[0073]FIG. 10 shows a graphical display of utility intervals;

[0074]
[0074]FIG. 11 shows a graphical display of attributes;

[0075]
[0075]FIG. 12 shows a graphical display which portrays the ranking of four motorcycles.
DETAILED DESCRIPTION OF THE INVENTION
Handling Oualitative and Quantitative Information

[0076]
The present invention permits the assessment of both quantitative and qualitative information which are subject to a range of uncertainties. Instead of using a decision matrix, the present invention describes a decision problem using a generalised decision matrix, an example of which is shown in Table 2 for the car selection problem which was described above in relation to Table 1. The main difference between a decision matrix and a generalised decision matrix is that the element of the latter can be a value or a distribution in a belief structure to accommodate uncertainties in human judgements.
TABLE 2 


Generalised Decision Matrix 
  Fuel    
 Engine  Consumption    General 
 Size  (mile/gallon)  Price  . . .  Quality 
 
Car 1  1400 cc  35 (50%)  £8,000  . . .  Good (75%) 
  40 (50%)    Excellent (25%) 
Car 2  1500 cc  40 (33%),  £9,000  . . .  Good (35%) 
  45 (33%) 
  50 (33%)    Excellent (65%) 
. . .  . . .  . . .  . . .  . . .  . . . 
Car N  1100  45 (25%),  £7,000  . . .  Average (15%) 
  46 (25%) 
  48 (25%),    Good (70%) 
  49 (25%) 
     Excellent (5%) 
     Unknown (10%) 


[0077]
Decision making with multiple criteria is based on the assessment of criteria. For instance, the quality of a motor engine may not be properly assessed without taking into account relevant quality criteria such as quietness, responsiveness, fuel consumption, vibration and starting, as shown in FIG. 1. Similar to the motor engine example, any general (upperlevel) criterion of an object can be assessed through its basic (lowerlevel) criteria, as shown in FIG. 2 or through a multilevel hierarchy of criteria.

[0078]
Due to the subjective nature of the criterion, the quality of a motor engine can be expressed in the present invention using certain standards in terms of linguistic evaluation grades such as poor, indifferent, average, good and excellent. For example, the quality of an engine may be described using the following distribution,

S(Quality)={(poor, 0.01)}, (indifferent, 0.1), (average, 0.15), (good, 0.63), (excellent, 0.05)} (1)

[0079]
which reads that the quality of the engine is 1% poor, 14% indifferent, 15% average, 63% good, and 5% excellent. The distribution provides a panoramic view of the engine's quality as far as the quality criteria are concerned. The percentages in equation (1) are referred to as the degrees of believe to which the engine's quality is assessed to individual grades. For example, 63% good means that the quality of the engine is assessed to the grade “good” to a degree of 63%.

[0080]
An assessment of quality is normally generated by aggregating more than one quality criterion. The quality criteria could be either quantitative or qualitative, and can be assessed in different ways. For instance,fuel consumption is a quantitative criterion and could be assessed using a quantity such as how many miles a motor vehicles can travel per gallon of fuel (mpg). On the other hand, it is more natural to assess a qualitative criterion using a set of grades appropriate for this criterion but not necessarily the same set as that used for assessing other criteria. In terms of quietness, for example, it is natural to judge that an engine is very quiet, quiet, normal, noisy or very noisy; in terms of vibration, it is common to judge that an engine vibrates heavily, normally or lightly.

[0081]
To aggregate both quantitative and qualitative criteria, the relationships amongst various sets of grades have to be properly interpreted. For instance, the performance of a motor engine is said to be good if it is quiet, its responsiveness is good, its fuel consumption is low (39 mpg for example), its vibration is normal, and its starting is also good. In the above aggregation, it is implied that a quiet engine means that the quality of the engine is good as far as quietness is concerned. In other words, the grade quiet in the quietness assessment is equivalent to the grade good in quality assessment. Similarly, in the above aggregation if the fuel consumption of an engine is 39 mpg then its quality is judged to be good as far as fuel consumption is concerned.

[0082]
In general, if both quantitative and qualitative criteria are included in a decision making problem it is necessary to transform various sets of assessment grades to a consistent framework so that they can be compared and aggregated consistently. In the following sections, techniques are exemplified which facilitate the transformation.
RuleBased Quantitative Information Transformation

[0083]
As discussed in the previous section, different linguistic evaluation grades may be used to describe the same standard. The equivalence between an evaluation grade and its corresponding standard can be established using equivalence rules to transform various sets of grades to a unified set. To transform quietness assessment to quality assessment, for example, the following simple equivalence rule could be established.

[0084]
Suppose an evaluation grade “very noisy” in a quietness assessment is equivalent to a grade “poor” in a quality assessment, “noisy” equivalent to “indifferent”, “normal” to “average”, “quiet” to “good”, and “very quiet” to “excellent”. Then one could say that the set of grades {very noisy, noisy, normal, quiet, very quiet} in quietness assessment is equivalent to the set {poor, indifferent, average, good, excellent} in quality assessment.

[0085]
The above equivalence is based on the fact that individual grades in the two sets are judged to be equivalent on the onetoone basis. In the case of transforming vibration assessment (heavily, normally or lightly) to quality assessment, however, the grade “heavily” for vibration criterion may imply a “poor” grade of engine quality to a degree of 80% and an “indifferent” grade to 20%. In general, a grade for a basic criterion may imply several grades for a general criterion to certain degrees. Suppose:

[0086]
H_{n }is the n^{th }grade for assessment of a general criterion

[0087]
K_{m,i }is the m^{th }grade for assessment of the l^{th }basic criterion

[0088]
α_{n,m }is the degree to which K_{m,i }implies H_{n }

[0089]
γ_{m,i }is the degree to which the l^{th }basic criterion is assessed to K_{m,i }

[0090]
β_{n,l }is the degree to which the l^{th }basic criterion is assessed to K_{m,i }

[0091]
Then, an assessment {(K
_{m,l},γ
_{m,l}),m=1, . . . M} ml under a set of grades {K
_{m,i}m=1, . . . M,} can be equivalently transformed to an assessment {(H
_{n},β
_{n,l}),n=1, . . . ,N} under another set of grades {H
_{n},n=1, . . . ,N} using the following matrix equation
$\begin{array}{cc}\left[\begin{array}{c}{\beta}_{1,l}\\ {\beta}_{2,l}\\ \vdots \\ {\beta}_{N,l}\end{array}\right]=\left[\begin{array}{cccc}{\alpha}_{1,1}& {\alpha}_{1,2}& \dots & {\alpha}_{1,M}\\ {\alpha}_{2,1}& {\alpha}_{2,2}& \dots & {\alpha}_{2,M}\\ \vdots & \vdots & \u22f0& \vdots \\ {\alpha}_{N,1}& {\alpha}_{N,2}& \dots & {\alpha}_{N,M}\end{array}\right]\ue8a0\left[\begin{array}{c}{\gamma}_{1,l}\\ {\gamma}_{2,l}\\ \vdots \\ {\gamma}_{M,l}\end{array}\right]& \left(2\right)\end{array}$

[0092]
The values of α
_{n,m}(m=1, . . . ,M and n=1, . . . ,N) should satisfy the equation
$\sum _{n=1}^{N}\ue89e{\alpha}_{n,m}=1$

[0093]
and are determined by the following rules extracted from decision makers:

[0094]
A grade K_{m,l }implies

[0095]
a grade H_{1 }to a degree of α_{1,m};

[0096]
a grade H_{2 }to a degree of α_{2,m}; . . . ; and

[0097]
a grade H_{N }to a degree of α_{N,m }with m=1, . . . , M.

[0098]
Because the values of α_{n,m}(m=1, . . . ,M and n=1, . . . ,N) are determined using rules, this approach can be termed a rulebased information transformation technique.
UtilityBased Information Transformation

[0099]
In the transformation technique described in the previous section it was assumed that the original assessment is equivalent to the transformed assessment in terms of value (also called utility) to decision makers, though the utilities of both assessments were not known explicitly. The utility of an assessment is given by the weighted sum of the utilities of grades using the degrees of belief as weights. The utility of a grade is a real number that is normally between 0 (the value for the most unfavourable grade) and 1 (the value for the most favourable grade). The utility of a grade represents a value of the grade to the decision maker. It is used to measure the decision maker's preferences towards a grade. Therefore, there is an element of subjectivity in utility estimation.

[0100]
Suppose the utilities of all grades are already given by a decision maker for both sets of grades {K
_{m,l},m=1, . . . ,M} and {H
_{n},n=1, . . . ,N}, denoted by u(K
_{m,l}) and u(H
_{n}). Then, an assessment {(K
_{m,l}γ
_{m,l}),m=1, . . . ,M} under the set of grades {K
^{m,i},m=1, . . . ,M} can be equivalently transformed to another assessment {(H
_{n},β
_{n,l}),n=1, . . . ,N} under the set of grades {H
_{n},n=1, . . . ,N} using the following matrix equation
$\left[\begin{array}{c}{\beta}_{1,l}\\ {\beta}_{2,l}\\ \vdots \\ {\beta}_{N,l}\end{array}\right]=\left[\begin{array}{cccc}{\alpha}_{1,1}& {\alpha}_{1,2}& \dots & {\alpha}_{1,M}\\ {\alpha}_{2,1}& {\alpha}_{2,2}& \dots & {\alpha}_{2,M}\\ \vdots & \vdots & \u22f0& \vdots \\ {\alpha}_{N,1}& {\alpha}_{N,2}& \dots & {\alpha}_{N,M}\end{array}\right]\ue8a0\left[\begin{array}{c}{\gamma}_{1,l}\\ {\gamma}_{2,l}\\ \vdots \\ {\gamma}_{M,l}\end{array}\right]$ $\mathrm{where}$ ${\alpha}_{n,m}=\frac{u\ue8a0\left({H}_{n+1}\right)u\ue8a0\left({K}_{m,l}\right)}{u\ue8a0\left({H}_{n+1}\right)u\ue8a0\left({H}_{n}\right)},{\alpha}_{n+1,m}=1{\alpha}_{n,m},$
α
^{i,m}=0 (
i=1
, . . . ,N,i≠n,n+1) if
u(
H _{n})≦
u(
K _{m,l})≦
u(
H _{n+1}) for
n=1, . . .
,N−1
; m=1, . . .
,M (3)

[0101]
If the utilities of both sets of grades are not given then they can be determined using the following equal distance scaling equations:
$\begin{array}{cc}u\ue8a0\left({H}_{n}\right)=\frac{n1}{N1}\ue89e\text{\hspace{1em}}\ue89e\mathrm{for}\ue89e\text{\hspace{1em}}\ue89en=1,\dots \ue89e\text{\hspace{1em}},N\ue89e\text{\hspace{1em}}\ue89e\mathrm{if}\ue89e\text{\hspace{1em}}\ue89e{H}_{n+1}\ue89e\text{\hspace{1em}}\ue89e\mathrm{is}\ue89e\text{\hspace{1em}}\ue89e\mathrm{preferred}\ue89e\text{\hspace{1em}}\ue89e\mathrm{to}\ue89e\text{\hspace{1em}}\ue89e\text{\hspace{1em}}\ue89e{H}_{n}& \left(4\right)\\ u\ue8a0\left({K}_{m,l}\right)=\frac{m1}{M1}\ue89e\text{\hspace{1em}}\ue89e\mathrm{for}\ue89e\text{\hspace{1em}}\ue89em=1,\dots \ue89e\text{\hspace{1em}},M\ue89e\text{\hspace{1em}}\ue89e\mathrm{if}\ue89e\text{\hspace{1em}}\ue89e{K}_{m+1,l}\ue89e\text{\hspace{1em}}\ue89e\mathrm{is}\ue89e\text{\hspace{1em}}\ue89e\mathrm{preferred}\ue89e\text{\hspace{1em}}\ue89e\mathrm{to}\ue89e\text{\hspace{1em}}\ue89e{K}_{m,l}& \left(5\right)\end{array}$
Quantitative Data Transformation

[0102]
A quantitative criterion is assessed using numerical values initially. To aggregate a quantitative criterion together with other qualitative criteria, equivalence rules are extracted to transform a value to an equivalent distribution using belief degrees on the chosen set of grades. For instance, a fuel consumption of 50 mpg of a motor engine may mean that the quality of the engine is “excellent” as far as fuel consumption is concerned. In other words, the 50 mpg fuel consumption is equivalent to “excellent” engine quality as far as fuel consumption is concerned. Similarly, fuel consumptions of 44, 38, 32 and 25 mpg may be equivalent to “good” “average”, “indifferent” and “poor”, respectively. Any other numbers between 25 and 50 mpg can be made to be equivalent to a few grades with different degrees of belief. For example,fuel consumption of 42 mpg might be held to be equivalent to “good” to a degree of belief of 67% and “average” to a degree of belief of 33%.

[0103]
In general, to assess a quantitative criterion, for example, the l^{th }criterion, a set of grades K_{m,l}(m=1, . . . , M) can be chosen which suits the criterion best, and then the assessment based on this set of grades transformed to an assessment based on another set of grades, for example the set of grades used to assess the general criterion associated with the quantitative criterion.

[0104]
For any set of grades K_{m,l}(m=1, . . . ,M), a set of corresponding values k_{m}(m=1, . . . ,M) can be found that the quantitative criterion may take so that k_{m }is equivalent to the grade K_{m,l}, or

k_{m }means K_{m}(m=1, . . . , M). (6)

[0105]
Without losing generality, one can even define a set of numerical grades K
_{m,l}(m=1, . . . ,M) so that K
_{m,l}=k
_{m}(m=1, . . . ,M), with K
_{M,l }being the most favourable feasible value of the criterion and K
_{1,l }the least. Suppose this is the case. Then any value k
_{j}(K
_{m,l}≦k
_{m+1,l} ,m=1, . . . ,M−1) of the l
^{th }criterion can be expressed as:
$\begin{array}{cc}{k}_{j}=\sum _{m=1}^{M}\ue89e\left({K}_{m,l}\ue89e{s}_{m,j}\right)& \left(7\right)\end{array}$

[0106]
where
$\begin{array}{cc}{s}_{m,j}=\frac{{K}_{m+1,l}{k}_{j}}{{K}_{m+1,l}{K}_{m,l}},\begin{array}{cc}\text{\hspace{1em}}& {s}_{m+1,j}=1{s}_{m,j}\end{array}\ue89e\text{}\ue89e\mathrm{if}\ue89e\text{\hspace{1em}}\ue89e{K}_{m,l}\le {k}_{j}\le {K}_{m+1,l},\begin{array}{cc}\text{\hspace{1em}}& m=1,\dots \ue89e\text{\hspace{1em}},M1\end{array}& \left(8\right)\end{array}$

[0107]
s_{i,j}=0 for i=1, . . . M; i≠m,m+1

[0108]
The assessment of k_{j }in terms of the set of grades K_{m,l}(m=1, . . . , M) can be expressed as

{(K_{m,l} ,γ _{m,l}), m=1, . . . M} with

γ_{m,l} =s _{m,j}(m=1, . . . , M) (9)

[0109]
In many decision situations, a quantitative criterion may be a random variable and take several values with different probabilities. Such assessment information can be expressed using a random number: {(k
_{j},p
_{j}),j=1, . . . ,P} where k
_{j}(j=1, . . . ,P) are possible values that the criterion may take, p
_{j }is the probability that the criterion may take a particular value k
_{j }and P is the number of possible values that the criterion may take. Using equations (7) and (8), s
_{m,j}(m=1, . . . ,M) can be calculated for each k
_{j}(j=1, . . . ,P). The random number {(k
_{j},p
_{j}),=1, . . . ,P} can then be transformed to an assessment {(K
_{m,l},γ
_{m,l})m=1, . . . ,,M} under the set of grades {K
_{m,l},m=1, . . . ,M} using the following matrix equation
$\begin{array}{cc}\left[\begin{array}{c}{\gamma}_{1,l}\\ {\gamma}_{2,l}\\ \vdots \\ {\gamma}_{M,l}\end{array}\right]=\left[\begin{array}{cccc}{s}_{1,1}& {s}_{1,2}& \dots & {s}_{1,P}\\ {s}_{1,1}& {s}_{1,2}& \dots & {s}_{1,P}\\ \vdots & \vdots & \u22f0& \vdots \\ {s}_{M,1}& {s}_{M,2}& \dots & {s}_{M,P}\end{array}\right]\ue8a0\left[\begin{array}{c}{p}_{1}\\ {p}_{2}\\ \vdots \\ {p}_{P}\end{array}\right]& \left(10\right)\end{array}$

[0110]
When the quantitative criterion takes a deterministic number, such as k
_{j}, then p
_{j}=1 and p
_{i}=0 for i=1, . . . ,M and i≠j in equation (10). That is, for a deterministic value, equation (10) becomes
$\begin{array}{cc}\left[\begin{array}{c}{\gamma}_{1,l}\\ {\gamma}_{2,l}\\ \vdots \\ {\gamma}_{M,l}\end{array}\right]=\left[\begin{array}{c}{s}_{1,j}\\ {s}_{2,j}\\ \vdots \\ {s}_{M,j}\end{array}\right]& \left(11\right)\end{array}$

[0111]
This is equivalent to equation (9). This is the special feature of a deterministic criterion, and the analysis conforms to the previous analysis.

[0112]
In order to aggregate the basic quantitative criterion with other basic criteria, it is necessary to transform the assessment results {(K
_{m,l},γ
_{m,l})m=1, . . . ,M} under the set of grades {K
_{m,l}, m=1, . . . , M to {(H
_{n},β
_{n,l}),n=1, . . . ,N} under the set of grades {H
_{n},n=1, . . . ,N} of the general criterion. We can use equation (2) to perform the transformation. Combining equation (2) with equation (10), we can transform a deterministic number k
_{j }or a random number {(k
_{j},p
_{j}),j=1, . . . ,P} to {(H
_{n},β
_{n,1}), n=1, . . . ,N} using the following equation:
$\begin{array}{cc}\left[\begin{array}{c}{\beta}_{1,l}\\ {\beta}_{2,l}\\ \vdots \\ {\beta}_{N,l}\end{array}\right]=\left[\begin{array}{cccc}{\alpha}_{1,1}& {\alpha}_{1,2}& \dots & {\alpha}_{1,M}\\ {\alpha}_{2,1}& {\alpha}_{2,2}& \dots & {\alpha}_{2,M}\\ \vdots & \vdots & \u22f0& \vdots \\ {\alpha}_{N,1}& {\alpha}_{N,2}& \dots & {\alpha}_{N,M}\end{array}\right]\ue8a0\left[\begin{array}{cccc}{s}_{1,1}& {s}_{1,2}& \dots & {s}_{1,P}\\ {s}_{2,1}& {s}_{2,2}& \dots & {s}_{2,P}\\ \vdots & \vdots & \u22f0& \vdots \\ {s}_{M,1}& {s}_{M,2}& \dots & {s}_{M,P}\end{array}\right]\ue8a0\left[\begin{array}{c}{p}_{1}\\ {p}_{2}\\ \vdots \\ {p}_{P}\end{array}\right]& \left(12\right)\end{array}$
Dealing with Incomplete Information

[0113]
Following appropriate transformations, all criteria can be described in the same framework. Using the techniques described in the previous sections, it is possible to do so even if the criteria comprise quantitative and qualitative criteria, and if the quantitative criteria take random or precise numbers. An example of such an instance is the assessment of quality criteria of a motor engine using the following distributions under the same set of grades.

[0114]
S[quietness]={[good, 0.5], [excellent, 0.3]}

[0115]
S[responsiveness]={[good, 1.0]}

[0116]
S[fuel economy]=([indifferent, 0.5], [average, 0.5]}

[0117]
S[vibration]={[good, 0.5], [excellent, 0.5]}

[0118]
S[starting]={[good, 1.0]}

[0119]
In an ideal situation, the quality of an engine will be regarded as good if its responsiveness, fuel economy, quietness, vibration and starting are all assessed to be exactly good. However, such consensus assessments are rare, and criteria are often assessed to different evaluation grades, as shown in the above example. A further problem is that an assessment may not be complete. For example, the assessment for quietness is not complete as the total degree of belief in the assessment is 0.5+0.3=0.8. In other words, 20% of the belief degrees in the assessment are missing.

[0120]
To judge the quality of an engine and compare it with other engines, a question is how to generate a quality assessment for the engine by aggregating the various assessments of the quality criteria as given above, which could be incomplete. This question is common to most MCDA problems. The present invention provides a systematic and rational way of dealing with the aggregation problem.
Generate Basic Probability Masses

[0121]
In the engine quality assessment problem, each quality criterion plays a part in the assessment but no single criterion dominates the assessment. In other words, the quality criteria are of relative importance. This is true of any MCDA problem.

[0122]
Weights for each of the basic criteria W_{i}(i=1, . . . ,L) should reflect the relative importance of each basic criterion to the general criterion. They can be assigned using a few methods such as a method based on pairwise comparisons or simply according to the decision maker's judgement. For example, if the decision maker thinks that basic criterion 1 is twice as important as basic criterion 2, then he or she may assign {W_{1}=10,W_{2}=5} or {W_{1}=2,W_{2}=1} as long as they are relatively correct.

[0123]
The weights assigned by the decision maker need to be normalised to arrive at a set of normalised weights ω
_{i}(i=1, . . . ,L) using the following equation
$\begin{array}{cc}{\omega}_{i}=\frac{{W}_{i}}{\sum _{j=1}^{L}\ue89e{W}_{j}}\ue89e\text{\hspace{1em}}\ue89e\left(i=1,\dots \ue89e\text{\hspace{1em}},L\right);& \left(13\right)\end{array}$

[0124]
so that 0≦ω
_{i}≦1 and
$\begin{array}{cc}\sum _{i=1}^{L}\ue89e{\omega}_{i}=1& \left(14\right)\end{array}$

[0125]
The present invention uses a new evidential reasoning algorithm for criteria aggregation, which operates on probability masses as described in the following sections. Since criteria are of relative importance, the assessment of one criterion to a grade to certain degree does not necessarily mean that all criteria would be assessed to the grade to the same degree. For instance, if the quietness of an engine is assessed to be good to a degree of 50%, the quality of the engine would not necessarily be assessed to be good to the same degree. This is because the engine quality is also determined by the other four quality criteria.

[0126]
In the present invention, the definition of a basic probability mass takes into account the relative importance of criteria. Let β_{n,i }denote a degree of belief that the i^{th }basic criterion is assessed to a grade H_{n}. Let m_{n,i }be a basic probability mass representing the degree to which the i^{th }basic criterion (quality criterion) supports the hypothesis that the general criterion (quality) is assessed to the n^{th }grade H_{n}. Let H denote the whole set of grades, or H={H_{n},n=1, . . ., N}. m_{n,i }is then defined as the weighted degree of belief or probability mass as shown in equation (15). Let m_{H,i }be a remaining probability mass unassigned to any individual grade after the i^{th }basic criterion has been assessed. Then, m_{n,i }and m_{H,i }are given by

m _{n,i}=ω
_{i}β
_{n,i} n=1, . .
,N (15)
$\begin{array}{cc}{m}_{H,i}=1\sum _{n=1}^{N}\ue89e{m}_{n,i}=1{\omega}_{i}\ue89e\sum _{n=1}^{N}\ue89e{\beta}_{n,i}& \left(16\right)\end{array}$

[0127]
m_{H,i }is decomposed into two parts: {overscore (m)}_{H,i }and {tilde over (m)}_{H,i }with m_{H,i}={overscore (m)}_{H,i}+{tilde over (m)}_{H,i},

[0128]
where
$\begin{array}{cc}\begin{array}{ccc}{\stackrel{\_}{m}}_{H,i}=1{\omega}_{i}& \mathrm{and}& {\stackrel{~}{m}}_{H,i}={\omega}_{i}\ue8a0\left(1\sum _{n=1}^{N}\ue89e{\beta}_{n,i}\right)\end{array}& \left(17\right)\end{array}$

[0129]
{overscore (m)}_{H,i }is the first part of the remaining probability mass that is not yet assigned to individual grades due to the fact that criterion i only plays one part in the assessment relative to its weight. In other words, {overscore (m)}_{H,i }provides the scope where other criteria can play a role in the assessment. {overscore (m)}_{H,i }should eventually be assigned to individual grades in a way that is dependent upon how all criteria are assessed and weighted. {tilde over (m)}_{H,i }is the second part of the remaining probability mass that is not assigned to individual grades due to the incompleteness in an assessment. {tilde over (m)}_{H,i }is proportional to ω_{i }and will cause the subsequent assessments to be incomplete.
Combine Probability Masses

[0130]
In equations (15) to (17), the contribution of the i^{th }basic criterion to the assessment of the general criterion (see FIG. 2) is represented as the basic probability masses. A new algorithm is developed to aggregate the basic probability masses. An important feature of the algorithm is its capacity of handling incomplete assessments by providing the upper and lower bounds of probability masses, based upon which the range of combined assessments can be generated, as discussed in the following sections.

[0131]
Let I_{n,i}(n=1,2, . . . ,N), {overscore (I)}_{H,i }and {overscore (I)}_{H,i }denote the combined probability masses generated by aggregating the first i criteria. The following algorithm can be used to combine the first i criteria with the (i+1)^{th }criterion in a recursive manner.

I_{n,l}=m_{n,l }(n=1,2, . . . ,N) (18a)

I_{H,l}=m_{H,l} (18b)

Ĩ_{H,l}={tilde over (m)}_{H,l} (18c)

{overscore (I)}
_{H,l}={overscore (m)}
_{H,l} (18d)
$\begin{array}{cc}{K}_{i+1}={\left[1\sum _{t=1}^{N}\ue89e\text{\hspace{1em}}\ue89e\sum _{\underset{j\ne t}{j=1}}^{N}\ue89e{I}_{t,i}\ue89e{m}_{j,i+1}\right]}^{1}& \text{(18e)}\end{array}$

I _{n,i+1} =K _{i+1} [I _{n,i} m _{n,i+1} +I _{H,i} m _{n,i+1} +m _{H,i+1}](n=1, . . . ,N) (18i)

Ĩ_{H,i+1} K _{i+1} [Ĩ _{H,i} {tilde over (m)} _{H,i+1} +{overscore (I)} _{H,i i} {tilde over (m)} _{H,i+1} +Ĩ _{H,i} {overscore (m)} _{H,i+1}] (18g)

{overscore (I)}_{H,i+1} =K _{i+1} [{overscore (I)} _{H,i} +{overscore (m)} _{H,i+1}] (18h)

I _{H,i+1} ={overscore (I)} _{H,i+1} Ĩ _{H,i+1} (18i)

[0132]
i={1,2, . . . ,L−1}

[0133]
wherein I_{n,L }is the combined probability mass assigned to the n^{th }grade (n=1, . . . N),{overscore (I)}_{H,L }the combined probability that needs to be redistributed over the N grades, and Ĩ_{H,L }the remaining combined probability mass that is unable to be distributed to any specific grade due to insufficient information, and I_{H,L}={overscore (I)}_{H,L}+Ĩ_{H,L}.
Generate Combined Degrees of Belief

[0134]
After all L basic criteria have been aggregated, the overall combined probability masses are given by I
_{n,L}(n=1,2, . . . ,N), Ĩ
_{H,L }and {overscore (I)}
_{H,L}. {overscore (I)}
_{H,L }denotes the remaining probability mass that can be assigned to individual grades. In the present invention, it is assigned to all individual grades proportionally using the following normalisation process so as to generate the combined degrees of belief to the grade H
_{n}.
$\begin{array}{cc}\begin{array}{cc}{\beta}_{n}=\frac{{I}_{n,L}}{1{\stackrel{\_}{I}}_{H,L}}& n=1,2,\dots \ue89e\text{\hspace{1em}},N\end{array}& \left(\text{19a}\right)\end{array}$

[0135]
The degree of belief that is not assigned to any individual grades is assigned to the whole set H by
$\begin{array}{cc}{\beta}_{H}=\frac{{\stackrel{~}{I}}_{H,L}}{1{\stackrel{\_}{I}}_{H,L}}& \left(\text{19b}\right)\end{array}$

[0136]
It has been proven that the combined degrees of belief generated using the above normalisation process satisfy the common sense synthesis rules (CSSR)^{ } in MCDA whilst incompleteness in original assessments is preserved and represented by β_{H}. The generated assessment for a general criterion can be represented by a distribution {(H_{n},β_{n}),n=1, . . . ,N}, which reads that the general criterion is assessed to the grade H_{n }with the degree of belief β_{n}(n=1, . . . ,N).
Generate a Utility Interval

[0137]
There may be occasions where distributed descriptions are not directly comparable to show the difference between two assessments. In such circumstances, it is desirable to generate numerical values equivalent to the distributed assessments in some sense. The present invention introduces the concept of expected utility to define such a value. Suppose u(H
_{n}) is the utility of the grade H with u(H
_{n+1})>u(H
_{n}) if H
_{n+1 }is preferred to H
_{n}. If all assessments are complete and precise, then β
_{H}=0 and the expected utility which is calculated by
$u=\sum _{n=1}^{N}\ue89e{\beta}_{n}\ue89eu\ue8a0\left({H}_{n}\right)$

[0138]
can be used for ranking alternatives.

[0139]
Note that β_{H }given in Equation (19b) is the unassigned degree of belief representing the extent of the incompleteness (ignorance) in the overall assessment. Within the evaluation framework discussed in the previous sections, β_{n }provides the lower bound of the likelihood to which H_{n }is assessed to. The upper bound of the likelihood is given by a belief degree of (β_{n}+β_{H}). Thus the belief interval [β_{n}(β_{n}+β_{H})] provides the range of the likelihood that H_{n }may be assessed to. The interval will reduce to a point β_{n }if all basic assessments are complete (or β_{H}=0).

[0140]
If any basic assessment is incomplete, the likelihood to which H
_{n }may be assessed to is not unique and can be anything in the interval [β
_{n}, (β
_{n}+β
_{H})]. In such circumstances, three values are defined to characterise a distributed assessment, namely the minimum, maximum and average utilities. Without loss of generality, suppose H
_{1 }is the least preferred grade having the lowest utility and H
_{N }the most preferred grade having the highest utility. Then the maximum, minimum and average utilities are given by
$\begin{array}{cc}{u}_{\mathrm{max}}=\sum _{n=1}^{N1}\ue89e{\beta}_{n}\ue89eu\ue8a0\left({H}_{n}\right)+\left({\beta}_{N}+{\beta}_{H}\right)\ue89eu\ue8a0\left({H}_{N}\right)& \left(20\right)\\ {u}_{\mathrm{min}}=\left({\beta}_{1}+{\beta}_{H}\right)\ue89eu\ue8a0\left({H}_{1}\right)+\sum _{n=2}^{N}\ue89e{\beta}_{n}\ue89eu\ue8a0\left({H}_{n}\right)& \left(21\right)\\ {u}_{\mathrm{avg}}=\frac{{u}_{\mathrm{max}}+{u}_{\mathrm{min}}}{2}& \left(22\right)\end{array}$

[0141]
wherein u_{max}, u_{min }and u_{avg }are the best possible, worst possible and average performance indicators in terms of utility values respectively, and u(H_{n}) (n=1, . . . N) are the utility values of the grade H_{n}(n=1, . . . N) as mentioned previously.

[0142]
The present invention includes within its scope computer programs which perform the above described methods, carrier media storing said computer programs, and computers which are adapted to perform the above described methods. Typically, a computer would be adapted to perform the methods of the invention by virtue of running computer programs of the present invention. Suitable carrier media include, but are not limited to, hard discs, floppy discs, compact discs, tapes, DVD and memory devices such as PROMs and EEPROMs. Computer programs, such as an embodiment which is exemplified below, can allow users to enter the transformation rules, to define assessment grades, to conduct evidence mapping processes and to aggregate multiple criteria using the ER algorithm. Additionally, the computer program can provide a graphical display of the results of an assessment. Computer programs can be provided which interface with commercially available operating systems or specific programs. The skilled reader will readily appreciate how such interfacing can be achieved.
EXAMPLE

[0143]
Assessment Criteria

[0144]
In this example a motorcycle assessment problem is examined using both complete and incomplete (imprecise) data of both a quantitative and qualitative nature. The belief structure will be used to facilitate continuous and imprecise assessments for qualitative criteria. For quantitative criteria, both certain and random numbers are taken into account. The transformation techniques are used to transform the various types of information into a unified framework. Software is used to support the analysis. The main window of the display produced by the software is shown in FIG. 3 for the motorcycle selection problem.

[0145]
The assessment problem has seven main criteria: Price, Displacement, Range, Top speed, Engine quality; Operation system and General finish. The first four criteria are quantitative and are measured using the following different units: poundsterling, cc, miles and mph, respectively.

[0146]
The last three criteria are qualitative and difficult to measure directly. Therefore they are assessed through detailed subcriteria. For example, engine quality is assessed through responsiveness, fuel consumption, quietness, vibration and starting; general finish through quality of finish, seat comfort, headlight, mirrors and horns. Operation system can be assessed through handling, transmission and brakes, which however are still difficult to assess directly and therefore are evaluated through more detailed subsub criteria. For example, handling is assessed through steering, bumpy bends, manoeuvrability and top speed stability; transmission through clutch operation and gearbox operation; and brakes through stopping power, braking stability and feel at control.

[0147]
Input Information

[0148]
Table 4 describes the motorcycle assessment problem, which involves four candidate motorcycles for assessment based on 29 criteria of a hierarchy as described in the previous section. The input information includes the relative weights among groups of criteria and the assessment outcome of each motorcycle on every criterion. The relative weights of the same group of criteria are shown in the brackets. Outcomes include precise numbers, random numbers and subjective assessments.

[0149]
Price, Displacement, Range and Top speed are all assessed using precise numbers. For examples, the price, displacement, range and top speed of Honda are given by £6199, 998 cc, 170 miles and 160 mph, respectively. Fuel consumption varies in different weather and road conditions. For example,fuel consumption is assessed on four conditions: (1) winter & urban, (2) winter & suburb, (3) summer & urban and (4) summer & suburb as well as the frequencies that a motorcycle is used in these conditions. For example, the fuel consumption of Honda is 31 mpg, 35 mpg, 39 mpg and 43 mpg under these four conditions with the equal frequency of 25% recorded by {[31, 0.25], [35,0.25], [39, 0.25], [43, 0.25]}. Quantitative numbers can be transformed to qualitative assessments using the techniques described previously. FIG. 4 shows an interface for implementing the rulebased data transformation technique which is supported by the software.

[0150]
For simplicity, the qualitative criteria in this example are all assessed using the same five evaluation grades, which are defined as Poor (H_{1}), Indifferent (H_{2}), Average (H_{3}), Good (H_{4}) and Excellent (H_{5}) and abbreviated by P, I, A, G and E respectively. The overall assessment of a motorcycle is also based on this set of grades. For example, the responsiveness of Yamaha is assessed to 30% good, and 60% excellent, denoted by {[G, 0.3], [E, 0.6]}. If different sets of grades are used for lowerlevel criteria, the rule or utilitybased techniques described in the previous sections can be used to transform them to the same set of grades. FIG. 5 shows an interface for implementing the rulebased qualitative information transformation technique. The utilitybased information transformation techniques are implemented in the software by estimating the utilities of grades. FIG. 6 shows an interface to support utility estimation.

[0151]
Imprecise assessments are lightly shaded in Table 4 and data absence is also assumed, as shown by the shaded blank boxes. Some judgements and random numbers are incomplete in the sense that the total degree of belief in an assessment is not summed to unity. For example, the assessment of the responsiveness of Yamaha is {[G, 0.3], [E, 0.6]} where the total belief degree is (0.3+0.6)<1 (or 30%+60%<100%). The assessment for the fuel consumption of Yamaha is {[28, 0.25], [34, 0.25], [38, 0.25]} with the total belief degree of 0.75[or 75%], since thefuel consumption data in urban areas in winter are not available. All input information, either quantitative or qualitative, can be fed into software using its input dialogue windows such as those shown in FIGS. 7 and 8.
TABLE 4 


Assessment Data for Motorcycle Selection Problem 





[0152]
If traditional MCDA methods were applied to the above problem, then at best one would have to make efforts to try to find the missing information and eliminate the imprecision. This is assuming that such efforts are practical and cost effective. Otherwise, additional assumptions need to be made about these missing and imprecise assessments, or certain criteria have to be abandoned for further analysis. In either event, the outcome is less than satisfactory. In contrast, the present invention is well suited to solving the problem using the very information of Table 4. The software may be used to support the following analysis.

[0153]
Ranking and Results

[0154]
The present invention can operate on degrees of belief. To generate utility intervals, it is necessary to estimate the utilities of values and grades. The certain monetary equivalent (CME) approach can be used to estimate the utilities of quantitative criteria. Take price for example. Suppose for this range of motorcycles the highest acceptable price is “£9,000” and the lowest possible price is “£5,000”. Note that the price is a cost criteria and therefore low price is preferred. First of all, the utility of price is normalised by assigning u(9000)=0 and u(5000)=1.

[0155]
Following the procedure of the CME approach, a price value having the average utility of £9,000 and £5,000 is identified first. Suppose the price value is ±7,500. Thus u(7500)=(u(9000)+u(5000))/2=0.5. Furthermore, suppose £6,500 has the average utility of £7,500 and ±5,000, or u(6500)=(u(7500)+u(5000))/2=0.75, and £8,500 has the average utility of £9,000 and £7,500, or u(8500)=(u(9000)+u(7500))/2=0.25.Let K_{1,1}=9000, K_{2,1}=8500, K_{3,1}=7500, K_{4, 1}=6500, K_{5,1}=5000. Note that if no preference information is available then a linear marginal utility function could be assumed for price. Under this assumption there would be {overscore (K)}_{1,1}=9000, {overscore (K)}_{2,1}=8000, {overscore (K)}_{3,1}=7000, {overscore (K)}_{4,1}=6000, and {overscore (K)}_{5,1}=5000.

[0156]
The probability assignment approach could be used to estimate the utilities of the five evaluation grades for the qualitative attributes. To illustrate the process and simplify discussion, suppose the utility of the five evaluation grades are equidistantly distributed in the normalised utility space, or u(P)=0, u(I)=0.25, u(A)=0.5 u(G)=0.75, u(E)=1.

[0157]
In Table 4, the criteria are of a threelevel hierarchy. In the present example, each group of the bottom level criteria associated with the same upperlevel criterion are first aggregated to generate an assessment for the upperlevel criterion. Once the assessments for a group of upperlevel criteria associated with the same higherlevel criterion are all generated, these assessments can be further aggregated in the same fashion to generate an assessment for the higherlevel criterion. This hierarchical aggregation process is based on the techniques previously described herein, and implemented in the software. The assessment of each motorcycle on any criterion can be reported graphically in the software, as shown in FIGS. 9 and 10, which display data concerning the quality of the Honda engine. Table 5 shows the final assessments generated using the software for the four motorcycles by aggregating all the criteria shown in Table 4. The comparison and ranking of the four motorcycles on the overall criterion and other selected criteria can be reported graphically as shown in FIGS. 11 and 12.

[0158]
The above results show that Honda is clearly the most recommended motorcycle as its minimum utility is larger than the maximum utilities of the other motorcycles. This is logical as it has the best engine quality, excellent general finish and relatively low price. Yamaha is ranked the second due to its low price followed by Kawasaki. BMW is ranked the last due to its high price and below average transmission and handling system. The above ranking is conclusive for the weights provided despite the imprecision and absence of some data. This shows that decision could be made on the basis of incomplete information. Note, however, that the above ranking is the personal choice of the decision maker who provided the weights of all the criteria and also estimated their marginal utilities. This means that given the same assessment data shown in Table 4, another decision make may achieve a different ranking.
TABLE 5 


Overall Assessment of Motorcycles 
Assessment  Kawasaki  Yamaha  Honda  BMW 

Distributed  {[P, 0.0515],  {[P, 0.0455],  {[P, 0.003],  {[P, 0.0888], 
assessment  [I, 0.0858],  [I, 0.1153],  [I, 0.0782],  [I, 0.2499], 
 [A, 0.2147],  [A, 0.2106],  [A, 0.1646],  [A, 0.1904]. 
 [G, 0.5693],  [G, 0.3232],  [G, 0.5374],  [G, 0.1155], 
 [E, 0.0572]}  [F, O.2745]}  [E, 0.2010]}  [E, 0.3288]} 
Maximum  0.6345  0.6820  0.7217  0.5996 
utility 
Minimum  0.6130  0.6511  0.7059  0.5864 
utility 
Average  0.6237  0.6665  0.7138  0.5732 
utility 
Ranking  3  2  1  4 


[0159]
It will be appreciated that this assessment is for exemplary purposes only, and that the invention is not limited in its scope by the specific disclosures of the example.