Answers are grounded in this paper and cite their sources.
Engineering design decisions often require the generation and comparison of design alternatives in an environment of multiple design criteria. To aid in such decision-making, designers turn to mathematical formalisms to assist in ranking the alternatives. A very popular approach to ranking alternatives against a set of n criteria is the weighted sum method (WSM). Popularity of the WSM method derives mainly from its intuitively compelling form and ease of use. It commonly takes the form (1)Rj=∑i=1nwiqijwhere the rating of the jth alternative, Rj, is the sum of criteria (or attribute) measures, qij, times weights, wi > 0, applied to each measure. The overwhelming preponderance of the more recent literature treats wi as an importance measure of each criterion, and although not necessary, it is also common to impose the condition (2)∑i=1nwi=1Although this is a very simple approach to ranking alternatives, it is, unfortunately, subject to a set of conditions that, while recognized in earlier literature [1,2], are frequently neglected, thus rendering the resulting ranking invalid. The purpose of this note is to serve as a reminder that the WSM method is neither as simple nor as universally applicable as the more recent literature [3–6] would imply.To begin, note that Eq. (1) is a linear preference model on the n criteria, qi. Hence, its use is restricted to the case where the measures of qi are linear functions of the value provided by each criterion. It is easy to see that this restriction imposes a rather severe limitation on the use of this preference model. For example, an increment of food would prove significantly more valuable to a person who has not eaten in several days than the same increment would be to the same person had she just consumed a large meal. Indeed, for most criteria, the value of an incremental improvement diminishes as the criterion is more fully satisfied. Fortunately, we can remove this limitation by writing Eq. (1) in the more general form (3)Rj=∑i=1nwivi(qij)where vi(qij) are “values” associated with their respective criteria measures.Second, an underlying condition key to the validity of the WSM is that the values of the criteria are linearly independent. This means that the value attributed to each criterion must be independent of the values attributed to all other criteria. If this condition is not met, the WSM will not produce a valid ranking. A test for the linear independence of the criteria is given by Fishburn [7]. Suffice to say that linear independence is not the norm.Next, consider the meaning of a ranking index such as Rj. Clearly, if Ra > Rb, we would expect to prefer alternative a to alternative b. Likewise, if Rb > Ra, we would expect to prefer alternative b to alternative a. If Ra = Rb, we would expect to be indifferent between alternatives a and b. It is this indifference statement that determines a condition on the criteria weights, wi. If we are indifferent between alternatives a and b, then the following condition must apply: (4)Ra=∑i=1nwivi(qia)=∑i=1nwivi(qib)=Rbwhich is to say (5)∑i=1nwiΔvi,a−b=0where Δvi,a − b = vi(qia) − vi(qib). Consider the simple case where n = 2 (which is equivalent to the case where the two alternatives differ only by two criteria). Then, the equations (6)w1Δv1,a−b+w2Δv2,a−b=0and (7)w1+w2=1impose conditions on wi. Rewriting Eq. (6)(8)w1w2=−Δv2,a−bΔv1,a−bwhich is to say that the ratio of the weights is determined by the indifference ratio in a trade between value measures of criterion 1 and criterion 2. That is, it is dependent on the ratio of the amount of a that one would be willing to “pay” to gain a defined increment in b. In short, the weights define an indifference surface, and they do not denote the relative importance of the criteria.Differentiating Eq. (3) on an indifference surface (9)∑i=1nwi∂vi(qi)∂qidqi=0In the n = 2 case, Eq. (9) becomes (10)w1∂v1(q1)∂q1dq1+w2∂v2(q2)∂q2dq2=0Rearranging terms (11)ω1/2=w1w2=−(∂v2(q2)/∂q2)dq2(∂v1(q1)/∂q1)dq1=−(∂v2(q2))/∂q2(∂v1(q1))/∂q1(dq2dq1)for all values of q1 and q2. Note that the term dq2/dq1 is a preference determined by the decision maker for combinations of q1 and q2. Thus, Eq. (11) imposes a condition on the relationships between the vi(qi). Equation (11) also makes it clear that the simple form of Rj as given by Eq. (1) is valid only in the rare case that all the derivatives ∂vi(qi)/∂qi are constants with respect to qi such that we may set vi = qi for all i. This means that equal increments of qi have an equal value regardless of qi itself.But wait! There is yet another condition that must be met. The weights as defined by the WSM (Eq. (1)) are constants. In order to assure that they are indeed constant, it is necessary to impose conditions on the criteria value measures. Obviously, since the weights are constants, their ratios must also be constants. Thus, ω1/2 is a constant that defines the relationship between the value measures, vi(qi).Let us see how this works with a simple example. Consider two criteria (attributes) in an aircraft design, say, cruise speed and range. It is important that the criteria be observable and measurable. The measurable quantities of criteria i are qi. In our example, cruise speed, qs, can be measured in knots (kts) and range, qr, in nautical miles (nm). But only under special conditions could these measured values be the vi(qi) values. We are free to assign the functions vi(qi) only within certain constraints: (1) the vi(qi) must be linearly increasing with the value assigned to different qi measures and (2) if two different measures of qi, say qia and qib, have the same value, they must both be assigned the same numerical value of vi(qi). In short, amounts qi map into their perceived values vi so as to satisfy the linearity requirement of Eq. (3). Essentially, the mapping monetizes values of qi but not necessarily such that different qi are monetized to the same monetary units. The ratios, ω, accommodate both differences in units and preference. Preference functions are unique only up to an affine transformation. Hence, in order to map the function vi(qi), we may arbitrarily choose a base point, qi0, and a small increment, Δqi0, and assign arbitrary values to both qi0 and Δqi0, namely, vi0 and Δvi0. For example, we could choose qr0 = 500 nm and Δqr0 = 100 nm, and vr0 = 10 and Δvr0 = 1. We then step forward to qr1 = qr0 + Δqr0 and, at qr1, assess the magnitude of Δqr1 that will result in Δvr0 = Δvr1 = 1. This is determined by the value that the decision-maker assigns to increments in qr as a function of qr. We continue this process until we have mapped vr(qr) over the full range of qr. Then, when the value functions, vr(qr) and vs(qs), are fully determined, the weight ratio, ωr/s, can be determined using Eq. (11). Looking at Eq. (11), we see that the weight ratio scales both the units of vr and vs relative to each other and accommodates the decision-maker's preference for trades between these criteria. The astute reader will note that instead of arbitrarily choosing the value functions vi(qi) and using these functions to determine the weight ratios, one could instead arbitrarily select the weight ratios and use these ratios to determine the value functions. This further illustrates the point that the weights in Eq. (1) do not reflect the relative importance of the different criteria.If all of the above conditions are met, the WSM will give an accurate representation of a decision-maker's preference vis-à-vis a preference ranking of the alternatives. Conversely, if all of the above conditions are not met, the resulting ranking of alternatives will be meaningless. Therefore, we see that a correct implementation of the very simple rating scheme provided by Eq. (1) is not simple at all provided one wants it to yield useful results, and we note that the common incorrect use of this scheme is likely to provide seriously misleading results. Indeed, we often pay substantial penalties for our failure to adhere to rigorous mathematical procedures. A far more thorough treatment of this topic is provided by Abbas [8].
No takes yet. Share an insight, caveat, or question.
George A. Hazelrigg (2019) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: