betaFeedback

WorksMathematics

Ordering Is Not Deciding: Towards Preference-Consistent Multi-Objective Design–Decision Optimisation

Abstract

In integrated design and group decision-making, Pareto filtering identifies non-dominated alternatives but does not determine which design best fits actors' preferences. Applying those preferences after filtering does not guarantee that the best-fit solution will be found: the filter may already have excluded the alternative with the highest aggregated preference, even when all combinations of fixed performance-improvement directions are considered. We compare Pareto–based selection with Integrative Maximisation of Aggregated Preferences (IMAP) on constructed multi-objective design–decision problems, using identical feasible solution spaces and actor preferences for both methods. The problems involve two system performances and up to four preference functions elicited from up to two actors. They cover independent and interdependent relationships between feasible performance values, monotonic and non-monotonic preference functions, and conflicting preferences for the same performance. In some cases, Pareto filtering excludes the best-fit solution, whereas IMAP identifies it. These counterexamples show that Pareto optimality does not guarantee a preference consistent decision, even in problems much simpler than real-world design and group decision-making problems. Pareto analysis can support ordering and exploration in performance space. IMAP instead models and aggregates actor preferences before optimisation and maximises the resulting aggregate preference over the full feasible and acceptable solution space.

Keywords: Pareto dominance concept; Integrative Maximisation of Aggregated Preferences (IMAP); Multi-objective design optimisation (MODO); Multi-criteria decision-making (MCDM); Design–decision systems; Constraint Programming (CP); Mixed-Integer (Non)Linear Programming (MI(N)LP); Parametric Design; Socio-technical Algorithmics; Complex design–decision problems.

Introduction

Integrated design and group decision-making share a common task: identifying a feasible design alternative that best fits the interests of the actors involved. An alternative is a possible solution to the problem, described by system performances such as completion time and cost. These performances, also called objectives in many conventional optimisation approaches, allow alternatives to be compared but do not by themselves determine which alternative is best. That depends on the actors’ preferences concerning different performance levels. The alternative that best fits their aggregated interests is referred to here as the best fit for common purpose [17, 16].

An actor’s preference concerning a performance can be represented by a preference function that maps each feasible performance value to a degree of preference. In other words, a preference function represents how desirable an actor considers different levels of a performance. Integrating such functions into a design–decision problem requires Preference Function Modelling (PFM) [2]. Preference is subjective, relative, and context-dependent and is therefore measured on a one-dimensional affine scale that provides a basis for choice. Note that preference, utility, worth, and value are synonyms [3].

A preference function may be monotonically increasing or decreasing, but it need not be linear or even monotonic. It may represent acceptable or most-preferred ranges, thresholds, or saturation effects (e.g., system reliability or maintenance investments), where further improvements beyond an already highly preferred performance yield only marginal preference gains. Alternatively, it may be non-monotonic, with an interior maximum or minimum corresponding to a preferred or non-preferred target value (e.g., project completion time, urban density, staffing level, noise level, or operating temperature), where performance values on either side of the target level are respectively less preferred or more preferred than the performance of interest at that target level. Moreover, different actors may have conflicting preference functions concerning the same performance. In design–decision problems, preference conflicts can therefore arise not only between performances but also within a single performance [17, 14].

Thus, every multi-criteria design–decision method that ends in a choice relies on preference information, whether explicitly represented or implicitly assumed. If the applied preference information does not adequately represent the actors’ interests, the resulting choice may be decision-invalid. Methods differ in where and how preference information enters the design–decision process [10]. In Pareto-supported optimisation, a fixed improvement direction (minimisation or maximisation) is prescribed for each performance. Pareto filtering then orders the solution space by dominance to identify the non-dominated alternatives: those for which no feasible alternative improves at least one performance without worsening any other under the prescribed directions. These alternatives form a Pareto front [6, 9], depending on the min–max orientation. For example, with two performances, four orientations are possible: min/min, min/max, max/min, and max/max, specifying the prescribed improvement direction for each performance. As additional performances are introduced, the number of possible orientations grows exponentially. The alternatives on each front remain unranked. Identifying the best fit for common purpose therefore requires the designer or decision-maker to introduce additional actor-preference information in an a posteriori selection step.

Such an a posteriori selection step is often performed using a weighted-utility method, in which the performances are combined into a single function, for example through a weighted sum. However, this may produce non-unique and preference-inconsistent rankings [18]. In contrast, in preference-based a priori methods, actor preferences are modelled and aggregated using PFM before optimisation. The Integrative Maximisation of Aggregated Preferences (IMAP) optimisation method, combined with PFM, directly maximises the aggregated preference over the complete feasible and acceptable design–decision space [17, 16, 13]. The resulting alternative is the preference optimum and represents the preference-consistent best fit for common purpose. This distinction raises a fundamental question underlying Pareto-supported optimisation: whether an ordering constructed independently of the actors’ preferences can identify their best fit for common purpose, even when preference information is applied a posteriori.

Pareto’s Claim

The view that alternatives can be ordered independently of the actors’ preferences originates in Pareto’s claim, as quoted in Barzilai Barzilai [ [2], that “the entire theory of economic equilibrium is independent of the notions of (economic) utility”. Barzilai [ [2] argues that this claim is incorrect: Pareto’s development requires mathematical operations that are not supported by the ordinal utility scales on which the claim relies. Barzilai [ [3] discusses related errors in group decision theory.

Pareto dominance nevertheless remains central to multi-objective optimisation, including engineering design optimisation and evolutionary algorithms [9, 8, 4, 5]. In a posteriori approaches, the absence of actor-preference information during Pareto filtering is addressed by introducing it afterwards to select from the resulting Pareto front [10, 6]. Whether this subsequent use of preference information is sufficient to identify the preference optimum remains to be established and therefore constitutes the central research question of this paper.

Further Limitations of Pareto Analysis

The limitations of Pareto analysis extend beyond Pareto’s preference-independence claim. More broadly, French [ [7] situates formal multi-criteria decision-making within an iterative process encompassing sense-making and modelling, analysis and exploration, and interpretation and implementation. From this overarching perspective, Pareto analysis supports the numerical ordering and exploration of alternatives in performance space but does not itself provide the preference measurement and aggregation required to identify a unique, actionable, and preference-consistent design–decision outcome.

More specifically, studies in the PFM tradition [17, 16, 13] argue that a Pareto front is insufficient as a basis for design–decision support. Pareto filtering identifies a subset of non-dominated alternatives based on prescribed performance-improvement directions but does not itself represent the actors’ possibly nonlinear, non-monotonic, or conflicting preference functions. Designers and decision-makers ultimately seek a design–decision configuration that best fits the actors’ aggregated interests, rather than a subset of alternatives that are merely non-dominated in performance space. Preference-aware Pareto variants incorporate preferences into the selection of candidates but still aim to approximate a Pareto set [1]. Additional preference information therefore remains necessary to move from numerical performance-space ordering to identifying the best fit for common purpose. This limitation is one of the principal motivations underlying the development of IMAP [17, 16, 13].

A separate limitation identified in many-objective optimisation is that the proportion of non-dominated alternatives can increase with the number of objectives, thereby reducing the discriminative power of Pareto dominance [11]. The present study addresses a different unresolved question. Even if a preference-consistent ranking is applied a posteriori, can it identify the preference optimum from the alternatives retained by Pareto filtering? Answering this question requires establishing whether the Pareto-filtered set necessarily contains the preference optimum.

Aim and Contributions

This paper examines whether Pareto-based selection necessarily retains the preference optimum when actor preferences are non-monotonic, target-based, or conflicting. The central research question is: Does the union of the Pareto fronts obtained from all possible minimisation–maximisation orientations necessarily contain the preference optimum when actor-preference information is applied a posteriori?

To address this question, Pareto-based selection and IMAP are compared across six constructed multi-objective design–decision problems. Each problem involves two system performances, up to two actors, and up to four preference functions. Together, the problems cover rectangular and circular solution spaces, uncoupled and coupled feasible performance values, monotonic and non-monotonic preference functions, and conflicting actor preferences concerning the same performance.

The comparison is performed on a like-for-like basis. All six problems involve two performances and at most two actors. For each problem, both methods use the same feasible and acceptable design–decision space, the same actor-preference information, and the same preference aggregation. The difference lies solely in when the resulting aggregated-preference function is applied. In Pareto-based selection, alternatives are first filtered by dominance under all four possible minimisation–maximisation orientations defined by the two performances, after which the aggregated-preference function is applied a posteriori. In IMAP, the same aggregated-preference function is used a priori as the optimisation criterion over the complete feasible and acceptable design–decision space.

By holding the solution space, actor-preference information, and preference aggregation constant, this construction isolates the effect of Pareto filtering. The six problems are not intended to estimate how frequently Pareto-based selection excludes the preference optimum in practice, but to determine whether it provides a general guarantee that the preference optimum remains available for subsequent preference-consistent selection. The paper thereby provides a direct comparison between preference-consistent a posteriori selection from the union of all Pareto fronts and a priori preference-based optimisation over the complete design–decision space.

The paper is structured as follows. § 2 introduces the common problem formulation, presents the six constructed problems, and compares the Pareto-based and IMAP outcomes. § 3 discusses the findings, their implications for real-world design–decision problems, the role of preference-guided search, and the resulting design–decision conclusions.

Pareto-Based Selection versus IMAP for Identifying a Design–Decision Optimum

This section compares Pareto-based selection and IMAP across six constructed multi-objective design–decision problems. The six problems use canonical settings from multi-objective optimisation [6, 9], allowing Pareto dominance, Pareto fronts, and preference aggregation to be examined without unnecessary complexity and providing a transparent basis for comparing the two methods. Both methods are applied to the same feasible and acceptable design–decision space using the same actor-preference information and preference aggregation. Because IMAP maximises the aggregated preference over the complete space by construction, it identifies the preference optimum. The comparison therefore examines whether the numerically constructed Pareto fronts, using the same aggregated-preference function a posteriori, also retain and identify this optimum. The preference optimum is labelled found when it is retained on at least one of the constructed Pareto fronts and selected from their union, and missed when it lies on none of these fronts and is therefore unavailable for subsequent selection. The software implementations of the six problems are available from the repository specified in the Data Availability statement.

Common Multi-Objective Design–Decision Problem Formulation

The IMAP–Pareto comparison is conducted entirely in the two-dimensional performance space. For all six problems, the number of performances is I=2I = 2 and the number of actors is at most K=2K = 2. This restriction enables all 2I=42^{I} = 4 Pareto orientations to be analysed exhaustively and the resulting preference landscapes, Pareto fronts, and preference optima to be visualised directly in two and three dimensions. The purpose of this simplified setting is analytical clarity rather than realism: it provides the minimal setting needed to examine whether Pareto-based selection guarantees retention of the preference optimum. A feasible and acceptable alternative is therefore represented directly by its performance vector f=(f1,f2)∈S\mathbf{f} = (f_{1}, f_{2}) \in S, where SS denotes the achievable set of feasible and acceptable performance vectors. The terms interior and boundary refer to this set. Each actor expresses a preference concerning each relevant performance through a preference function. The figures are computed from sampled performance vectors within each achievable set. The problems use monotonic and non-monotonic preference functions. A monotonic preference function, whether linear or non-linear, represents a consistent preference for more or less of a performance and therefore defines one fixed improvement direction over the feasible range. A non-monotonic preference function does not define one fixed improvement direction over this range. In particular, such a function may attain a global maximum at an interior performance value, with preference increasing towards this value from one direction and decreasing beyond it in the other.

Pareto criterion. An orientation e∈{−1,+1}I\mathbf{e} \in\{-1,+1\}^{I} fixes a minimisation or maximisation direction for each performance, where II is the number of performances, ei=+1e_i = +1 denotes maximisation, and ei=−1e_i = -1 denotes minimisation. Under orientation e\mathbf{e}, a performance combination a∈S\mathbf{a} \in S dominates another performance combination b∈S\mathbf{b} \in S if

a≻eb⟺eiai≥eibi for all iandejaj>ejbj for some j,(1)\mathbf{a} \succ_{\mathbf{e}} \mathbf{b} \Longleftrightarrow e_i a_i \ge e_i b_i \text{ for all } i \quad\text{and} \quad e_j a_j > e_j b_j \text{ for some } j, \tag*{(1)}

and the corresponding Pareto front is the set of performance combinations in SS that are not dominated under that orientation:

Pe(S)={a∈S:∄b∈S such that b≻ea}.(2)\mathcal{P}_{\mathbf{e}}(S) = \{\mathbf{a} \in S : \nexists\mathbf{b} \in S \text{ such that } \mathbf{b} \succ_{\mathbf{e}} \mathbf{a}\}. \tag*{(2)}

In general, II performances define 2I2^{I} possible orientations and corresponding Pareto fronts, although fronts obtained under different orientations need not be distinct. For the two performances considered here, the four orientations are min/min, min/max, max/min, and max/max. Pareto dominance and Pareto fronts are standard concepts [6, 10].

IMAP criterion. For KK actors and II performances, let zk,i(fi)z_{k,i}(f_i) denote the z-normalised preference of actor kk concerning performance ii. Since these preference functions are represented on a common linear preference scale, they can be consistently aggregated using the linear weighted-centroid operation performed by the A-FINE aggregator A\mathbf{A} [17]. The software implementation is available through the repository specified in the Data Availability statement. The resulting aggregated-preference function is

Z(f)=A(zk,i(fi);wk,i′)=∑i=1IAi(f)=∑i=1I∑k=1Kwk,i′zk,i(fi).(3)Z(\mathbf{f}) = \mathbf{A}(z_{k,i}(f_i); w'_{k,i}) = \sum_{i=1}^{I} A_i(f) = \sum_{i=1}^{I}\sum_{k=1}^{K} w'_{k,i}z_{k,i}(f_i). \tag*{(3)}

Here, Ai(f)=∑k=1Kwk,i′zk,i(fi)A_i(\mathbf{f}) = \sum_{k=1}^{K} w'_{k,i}z_{k,i}(f_i) is the weighted aggregation of the relevant actors’ preferences concerning performance ii and constitutes that performance’s contribution to Z(f)Z(\mathbf{f}). The effective contribution weight wk,i′=wkwk,iw'_{k,i} = w_k w_{k,i} combines the global weight wkw_k of actor kk with the actor’s local weight wk,iw_{k,i} for performance ii.

IMAP models and aggregates the actor preferences a priori and directly identifies the aggregated preference optimum

f∗=arg max⁡f∈S  Z(f).(4)\mathbf{f}^{*} = \underset{\mathbf{f}\in S}{\operatorname{arg\,max}}\; Z(\mathbf{f}). \tag*{(4)}

Thus, IMAP maximises the aggregated preference over the complete achievable set SS, including all alternatives on the Pareto fronts.

Like-for-like comparison. Both methods use the same achievable performance set SS, actor-preference functions, A-FINE aggregation, and resulting aggregated-preference function Z(f)Z(\mathbf{f}). In IMAP, the preference functions are aggregated a priori; for the like-for-like Pareto-based comparison, the same preference functions, weights, and A-FINE aggregation are applied a posteriori to the alternatives retained by Pareto filtering. For this comparison, IMAP, like Pareto-based selection, is formulated directly in performance space as P(f)P(\mathbf{f}), rather than through its standard design–decision-space formulation using the composite mapping P(f(x))P(\mathbf{f}(\mathbf{x})). IMAP maximises Z(f)Z(\mathbf{f}) over the complete set SS, including all alternatives on the Pareto fronts, whereas Pareto-based selection maximises Z(f)Z(\mathbf{f}) only over the union of the four fronts. The preference optimum is found if it lies on at least one front and is selected from their union; otherwise, it is missed. Furthermore, three structural properties guide the like-for-like comparison:

  • Separability. Because Z(f)Z(\mathbf{f}) is additively separable across performances, a change in fif_i affects Z(f)Z(\mathbf{f}) only through Ai(fi)A_i(f_i). When only one actor expresses a preference concerning performance ii, the shape of Ai(fi)A_i(f_i) is therefore determined by that actor’s preference function.

  • Direction. Pareto dominance requires a fixed minimisation or maximisation direction for each performance. A monotonic preference function provides such a direction, whereas a non-monotonic preference function with an interior global maximum does not define one fixed direction over the complete feasible range. To avoid relying on an arbitrarily prescribed orientation, all 2I2^I possible orientations are therefore considered.

  • Shape of SS. The geometry of the achievable set determines where the Pareto fronts lie. For the circular set considered here, the union of the four fronts contains the complete boundary. For the rectangular set considered here, each orientation yields one corner as its Pareto front; hence, the non-corner boundary points and all interior points lie on none of the four fronts.

Note on problem numbering and terminology. Each problem is identified by two numbers: the first denotes the problem number (1–6), and the second denotes the number of actors (1 or 2). Moreover, “peaked at” and “has a target value at” are used to mean “attains a global maximum at”. The “preference optimum” is the alternative that attains the maximum aggregated preference and is therefore the best fit for common purpose.

Uncoupled Rectangular Solution Space: Problems 1–2

For Problems 1–2, the solution space is the rectangle

S1={(f1,f2)∈R2:a≤f1≤b, c≤f2≤d}=[a,b]×[c,d].(5)S_1 = \left\{(f_1,f_2) \in\mathbb{R}^2 : a \le f_1 \le b,\ c \le f_2 \le d\right\} = [a,b] \times[c,d]. \tag*{(5)}

The performances are uncoupled: every feasible value of f1∈[a,b]f_1 \in[a,b] can be combined with every feasible value of f2∈[c,d]f_2 \in[c,d]. The four Pareto fronts reduce to the corner points (a,c)(a,c), (a,d)(a,d), (b,c)(b,c), and (b,d)(b,d) for the min/min, min/max, max/min, and max/max orientations, respectively. Under each orientation, the corresponding corner dominates every other point in S1S_1.

Rectangular solution space with corner points labelled by the four minimisation–maximisation orientations

Figure 1. Pareto “fronts” (corner points) under the four minimisation–maximisation orientations for the rectangular solution space S1S_1.

In Problems 1–2, one actor expresses preferences concerning both performances. With K=1K = 1 and I=2I = 2, the A-FINE aggregation in Eq. (3) simplifies to

Z(f1,f2)=A1(f1)+A2(f2)=w1z1(f1)+w2z2(f2).(6)Z(f_1, f_2) = A_1(f_1) + A_2(f_2) = w_1 z_1(f_1) + w_2 z_2(f_2). \tag*{(6)}

where zi(fi)z_i(f_i) is the actor’s z-normalised preference concerning performance ii, wiw_i is its effective contribution weight, and Ai(fi)=wizi(fi)A_i(f_i) = w_i z_i(f_i) is the corresponding contribution to Z(f1,f2)Z(f_1, f_2). The same formulation applies to the one-actor Problems 3–4.

Problem 1-1: P1, P2 both linear

Let the preference function for f1f_1 be linearly decreasing and that for f2f_2 linearly increasing, with equal local weights assigned to both performances. For one actor, each contribution Ai(fi)A_i(f_i) is a positive affine transformation of the corresponding preference function and is therefore strictly monotonic. Consequently, the preference optimum is f∗=(a,d)\mathbf{f}^* = (a,d), where f1f_1 attains its minimum and f2f_2 its maximum. This point coincides with the Pareto “front” of the min/max orientation, which follows directly from the two preference directions. Pareto-based selection therefore finds f∗\mathbf{f}^*, and both methods agree. This result follows analytically from the separable linear formulation and is independent of sampling.

Preference functions, aggregated-preference function, and Pareto front for Problem 1-1

Figure 2. Problem 1-1 with linear monotonic preferences and rectangular solution space S1S_1: (a,b) preference functions for f1f_1 and f2f_2; (c) aggregated-preference function ZZ, attaining its global maximum at the preference optimum f∗f^{*}; and (d) f∗f^{*} coinciding with the min/max Pareto “front” (corner point).

Problem 2-1: P1 linear & P2 non-monotonic with interior global max

The linearly decreasing preference function for f1f_1 is retained, while the preference function for f2f_2 is non-monotonic and attains its global maximum at an interior value tt, where c<t<dc < t < d. Consequently, A1A_1 is decreasing and A2A_2 attains its global maximum at tt, yielding the preference optimum f∗=(a,t)f^{*}=(a,t). This point lies on the left boundary of S1S_1, strictly between the corner points (a,c)(a,c) and (a,d)(a,d), and therefore lies on none of the four Pareto fronts. Because the preference function for f2f_2 does not define one fixed improvement direction over its complete feasible range, all four orientations are considered. Under each orientation, the corresponding corner dominates f∗f^{*} but has a lower aggregated preference. Pareto-based selection therefore misses f∗f^{*}, whereas IMAP identifies it.

Four-panel illustration of preference functions, aggregated-preference function, and Pareto fronts

Figure 3. Problem 2-1 with a linear decreasing preference for f1f_1, a non-monotonic preference for f2f_2 attaining an interior global maximum at tt, and rectangular solution space S1S_1: (a,b) preference functions for f1f_1 and f2f_2; (c) aggregated-preference function ZZ and preference optimum f∗f^*; and (d) the four Pareto “fronts” (corner points), the preferred performance levels, and f∗f^* on the left boundary of S1S_1 between two corner points.

Coupled Circular Solution Space: Problems 3–6

For Problems 3–6, the achievable set is a circle inscribed in the box [a,b]×[c,d][a,b] \times[c,d]. The corresponding solution space is

S2={f:(f1−m1h1)2+(f2−m2h2)2≤1}(7)S_2 = \left\{ \mathbf{f} : \left(\frac{f_1-m_1}{h_1}\right)^2 + \left(\frac{f_2-m_2}{h_2}\right)^2 \le1 \right\} \tag*{(7)}

where m1=a+b2m_1 = \frac{a+b}{2} and m2=c+d2m_2 = \frac{c+d}{2} are the centre coordinates, and h1=b−a2h_1 = \frac{b-a}{2} and h2=d−c2h_2 = \frac{d-c}{2} are the corresponding half-widths of the box. For the circular solution space considered here, h1=h2h_1 = h_2.

Pareto fronts under the four minimisation–maximisation orientations for the circular solution space $S_2$

Figure 4. Pareto fronts under the four minimisation–maximisation orientations for the circular solution space S2S_2. Each front is a boundary arc; together, the four fronts form the complete boundary of S2S_2.

Thus, the bounding intervals have equal widths in the depicted performance space. This circular representation facilitates transparent geometric comparison and visualisation; the physical quantities, values, and units represented by f1f_1 and f2f_2 need not be identical. The performances are coupled: the feasible values of f1f_1 and f2f_2 constrain one another, so not every combination within the bounding box is feasible. Each Pareto orientation defines one boundary arc. The min/min, min/max, max/min, and max/max fronts are the lower-left, upper-left, lower-right, and upper-right arcs, respectively. Together, the four fronts form the complete boundary of S2S_2, whereas the interior points lie on none of the four fronts. Problems 3–4 involve one actor and use the A-FINE aggregation in Eq. (6). Problems 5–6 involve two actors; accordingly, the A-FINE aggregation becomes

Z(f1,f2)=A1(f1)+A2(f2)=∑k=12wk,1′zk,1(f1)+∑k=12wk,2′zk,2(f2).(8)Z(f_1,f_2)=A_1(f_1)+A_2(f_2)=\sum_{k=1}^{2}w'_{k,1}z_{k,1}(f_1)+\sum_{k=1}^{2}w'_{k,2}z_{k,2}(f_2). \tag*{(8)}

Problem 3-1: P1P_1 linear & P2P_2 non-monotonic with interior global max

The linearly decreasing preference function for f1f_1 and the non-monotonic preference function for f2f_2 from Problem 2-1 are retained. Consequently, A1A_1 is decreasing, while A2A_2 attains its global maximum at the interior value tt. Because the performances are coupled in S2S_2, the preference optimum f∗f^* lies on the lower-left boundary arc, with f2f_2 close to but not equal to tt. Moving along this arc towards a lower value of f1f_1 requires f2f_2 to move away from tt, so f∗f^* balances the resulting preference gain in A1A_1 against the preference loss in A2A_2. This boundary arc is the min/min Pareto front. Because the non-monotonic preference function for f2f_2 does not define one fixed improvement direction, all four orientations are considered. Their union contains the complete boundary of S2S_2 and therefore retains f∗f^*. Pareto-based selection consequently finds f∗f^*.

Problem 3-1 preference functions, aggregated-preference function, and Pareto front

Figure 5. Problem 3-1 with a linear decreasing preference for f1f_1, a non-monotonic preference for f2f_2 attaining an interior global maximum at tt, and circular solution space S2S_2: (a,b)(a,b) preference functions for f1f_1 and f2f_2; (c) aggregated-preference function ZZ, attaining its global maximum at the preference optimum f∗f^*; and (d) f∗f^* located on the min/min Pareto front.

Problem 4-1: P1, P2 both non-monotonic with interior global max

Both preference functions are non-monotonic and attain their global maxima at interior values t1t_1 and t2t_2, with equal local weights assigned to the two performances. Consequently, A1A_1 and A2A_2 attain their global maxima at t1t_1 and t2t_2, respectively. Because (t1,t2)∈int⁡(S2)(t_1,t_2) \in\operatorname{int}(S_2), the preference optimum is f∗=(t1,t2)f^*=(t_1,t_2). This interior point lies on none of the four Pareto fronts and is dominated under each orientation by at least one boundary point. However, every point in the union of the fronts has a lower aggregated preference than f∗f^*. Paretobased selection therefore misses f∗f^*, whereas IMAP identifies it.

Four panels showing preference functions, aggregated-preference function, and Pareto fronts

Figure 6. Problem 4-1 with non-monotonic preferences for both f1f_1 and f2f_2, each attaining an interior global maximum, and circular solution space S2S_2: (a,b) preference functions for f1f_1 and f2f_2; (c) aggregated-preference function ZZ and preference optimum f∗f^*; and (d) the four Pareto fronts, the preferred performance levels, and f∗f^* in the interior of S2S_2.

Problem 5-2: P1,1P_{1,1}, P2,1P_{2,1} conflicting and linear & P1,2P_{1,2}, P2,2P_{2,2} equal non-monotonic with interior global max

Two actors express preferences concerning both performances, with equal local weights assigned to f1f_1 and f2f_2. Their preferences concerning f2f_2 are identical and non-monotonic, each attaining its global maximum at the same interior value t2t_2. Consequently, the aggregated contribution A2A_2 also attains its global maximum at t2t_2. The actors differ only in their preferences concerning f1f_1: P1,1P_{1,1} is linearly increasing, whereas P2,1P_{2,1} is linearly decreasing. Actor 1 has the larger global weight. The weighted aggregation of the opposing preferences therefore yields a linearly increasing contribution A1A_1. The global weights must be unequal for this problem: equal global weights would cause the opposing z-normalised linear preferences to cancel, making A1A_1 constant. The direction of A1A_1 is thus determined by the weighted aggregation of the actors’ conflicting preferences, rather than by either actor’s individual preference alone.

Because A1A_1 is increasing and A2A_2 attains an interior global maximum at t2t_2, the preference optimum f∗f^* lies on the upper-right boundary arc of S2S_2, with f2f_2 close to but not equal to t2t_2. Moving along this arc towards a higher value of f1f_1 requires f2f_2 to move away from t2t_2, so f∗f^* balances the preference gain in A1A_1 against the preference loss in A2A_2. This arc is the max/max Pareto front. Since the non-monotonic preferences concerning f2f_2 do not define one fixed improvement direction, all four orientations are considered. Their union retains f∗f^*, and Pareto-based selection therefore finds the same preference optimum as IMAP. If the actors’ global weights were interchanged, A1A_1 would instead be decreasing and the preference optimum would lie on the upper-left, min/max Pareto front.

Problem 6-2: P1,1P_{1,1}, P2,1P_{2,1} conflicting and non-linear & P1,2P_{1,2}, P2,2P_{2,2} equal non-monotonic with interior global max

Problems 5-2 and 6-2 form a controlled comparison: the actors’ preference directions remain unchanged, while only the shapes of their preference functions concerning f1f_1 change from linear to non-linear. Actor 1’s preference is increasing with diminishing returns, whereas Actor 2’s preference is decreasing at an increasing rate. The actors’ identical non-monotonic preferences concerning f2f_2 remain unchanged.

Although both preferences concerning f1f_1 are monotonic, their weighted aggregation A1A_1 is non-monotonic and attains its global maximum at an interior value f^1\hat{f}_1. The aggregated contribution A2A_2 likewise attains its global maximum at the interior value t2t_2. Because (f^1,t2)∈int⁡(S2)(\hat{f}_1,t_2) \in\operatorname{int}(S_2), the preference optimum is f∗=(f^1,t2)f^* = (\hat{f}_1,t_2).

The preference optimum lies on none of the four Pareto fronts and is dominated under each orientation by at least one boundary point. Nevertheless, every point in the union of the fronts has a lower aggregated preference than f∗f^*. Pareto-based selection therefore misses f∗f^*, whereas IMAP identifies it.

The comparison with Problem 5-2 isolates the effect of preference-function shape. Despite the unchanged individual preference directions, the aggregated contribution A1A_1 changes from monotonic to non-monotonic, and the preference optimum moves from the boundary to the interior of S2S_2. Consequently, the Pareto-based result changes from found in Problem 5-2 to missed in Problem 6-2.

Comparative Overview of the Six Problems

Table 1 summarises the location of the preference optimum and the outcome of Pareto-based selection for the six problems. Pareto-based selection finds the preference optimum in Problems 1-1, 3-1, and 5-2, where the optimum lies on at least one Pareto front. It misses the preference optimum in Problems 2-1, 4-1, and 6-2, where the optimum lies on none of the four fronts.

Problem - # of actors / types of PP-functionsIMAPPareto
S1S_1 - Rectangular
1-1 / P1P_1, P2P_2 both linearcorner pointfound (min/max front)
2-1 / P1P_1 linear & P2P_2 non-monotonic (interior max)edge pointmissed
S2S_2 - Circular
3-1 / P1P_1 linear & P2P_2 non-monotonic (interior max)boundary pointfound (min/min front)
4-1 / P1P_1, P2P_2 both non-monotonic (interior max)interior pointmissed
5-2 / P1,1P_{1,1}, P2,1P_{2,1} conflicting and linear & P1,2P_{1,2}, P2,2P_{2,2} equal non-monotonic (interior max)boundary pointfound (max/max front)
6-2 / P1,1P_{1,1}, P2,1P_{2,1} conflicting and non-linear & P1,2P_{1,2}, P2,2P_{2,2} equal non-monotonic (interior max)interior pointmissed

Table 1. Results for the six design–decision problems: the location of the IMAP preference optimum f∗f^* in SiS_i and whether Pareto-based selection recovers it as the design with the highest aggregated preference ZZ. The relevant front is shown in parentheses when f∗f^* is found; “missed” indicates that f∗f^* lies on none of the four possible Pareto fronts.

Problem 5-2 with two actors having opposing linear preferences for $f_1$, identical non-monotonic preferences for $f_2$, and circular solution space $S_2$

Figure 7. Problem 5-2 with two actors having opposing linear preferences for f1f_1, identical non-monotonic preferences for f2f_2, and circular solution space S2S_2: (a,c) the actors’ preference functions for f1f_1 and f2f_2; (b,d) the respective aggregated contributions A1A_1 and A2A_2, each attaining its global maximum at the indicated performance level; (e) aggregated-preference function ZZ and preference optimum f∗f^*; and (f) the four Pareto fronts, the preferred performance levels, and f∗f^* on the max/max front.

Six-panel figure showing preference functions, aggregated contributions, aggregated-preference function, and Pareto fronts

Figure 8. Problem 6-2 with two actors having opposing non-linear preferences for f1f_1, identical non-monotonic preferences for f2f_2, and circular solution space S2S_2: (a,c) the actors’ preference functions for f1f_1 and f2f_2; (b,d) the respective aggregated contributions A1A_1 and A2A_2, each attaining an interior global maximum; (e) the aggregated-preference function ZZ and preference optimum f∗f^*; and (f) the four Pareto fronts, the most-preferred performance levels, and f∗f^* in the interior of S2S_2.

In the rectangular solution space S1S_1, the corner optimum of Problem 1-1 coincides with the min/max front and is found, whereas the non-corner boundary optimum of Problem 2-1 lies on none of the four corner fronts and is missed. In the circular solution space S2S_2, the boundary optima of Problems 3-1 and 5-2 lie on the min/min and max/max fronts, respectively, and are found. The interior optima of Problems 4-1 and 6-2 lie on none of the four fronts and are missed. Problems 5-2 and 6-2 provide the controlled comparison. The individual preference directions are unchanged, but replacing the conflicting linear preferences concerning f1f_1 with conflicting non-linear preferences changes the aggregated contribution A1A_1 from monotonic to non-monotonic. Correspondingly, the preference optimum moves from the boundary in Problem 5-2 to the interior in Problem 6-2, and the Pareto-based outcome changes from found to missed.

Across these six simplified, canonical problems, Pareto-based selection misses the preference optimum when at least one actor has a non-monotonic preference function with an interior global maximum (Problems 2-1 and 4-1), or even when conflicting monotonic actor preferences concerning one performance aggregate into a non-monotonic preference function while the aggregated preference concerning the other performance is also non-monotonic (Problem 6-2). These findings demonstrate that Pareto-based selection does not, in general, guarantee a decision-valid multi-criteria outcome.

Conspection

The findings from the six constructed problems must be considered in the broader context of real-world design–decision optimisation. The failure of Pareto-based selection in three of these deliberately simplified, canonical problems does not provide an estimate of its failure rate in practice, but it disproves any general guarantee that Pareto filtering retains the preference optimum. If this guarantee already fails in a two-dimensional setting with at most two actors, four Pareto orientations, and explicitly defined preference functions, the substantially greater complexity of real-world design–decision problems provides no basis for assuming that it will be restored. The concluding subsection returns to the central research question.

Real-World Conditions Beyond the Constructed Problems

Real-world design–decision problems are substantially more complex than the constructed problems. Practical preference functions are seldom exclusively linear or monotonic; they may contain acceptable or most-preferred ranges, thresholds (e.g., budget caps and minimum availability requirements), saturation effects (e.g., system reliability and maintenance investments), and interior global maxima or minima (e.g., project completion time, urban density, staffing level, noise level, or operating temperature). Collectively, these characteristics demonstrate that the monotonic “less-is-better” and “more-is-better” preference representations implicitly assumed in Pareto analysis are often insufficient for representing real-world actor preferences. Multiple actors may also express heterogeneous or conflicting preferences concerning the same performance [17, 14]. Accordingly, one fixed minimisation or maximisation direction cannot generally represent the complete preference structure associated with a performance in design–decision optimisation.

In practical applications, performances are often functions of shared design–decision variables, f(x)f(\mathbf{x}), and are therefore commonly interdependent. Changing one design–decision variable may affect several performances simultaneously, producing coupled, non-linear, and constrained feasible spaces. The actors’ performance-to-preference functions must consequently be evaluated through the composite mapping P(f(x))P(f(\mathbf{x})) and aggregated across performances and actors. The resulting best-fit design–decision point is obtained by maximising the aggregated preference over the feasible and acceptable design–decision space. In general, this optimum need not coincide with the maximum of any individual preference function, because it emerges from maximising the aggregation of the preferences of multiple actors across multiple interdependent performances within the coupled design–decision space. Pareto dominance does not represent this performance-to-preference and preference-aggregation structure; it orders feasible performance combinations numerically under prescribed improvement directions and therefore cannot guarantee decision-valid outcomes.

Increasing dimensionality introduces two further limitations. For II performances, the number of possible minimisation–maximisation orientations grows exponentially as 2I2^{I}, making exhaustive construction of all corresponding Pareto fronts rapidly impractical. Simultaneously, the proportion of mutually non-dominated alternatives generally increases with the number of performances, reducing the discriminative power of Pareto dominance [11, 5]. Pareto analysis may therefore require an impractical number of orientations while producing increasingly large sets of non-dominated alternatives. Neither exhaustive orientation analysis nor enlarged non-dominated sets provide the preference measurement and aggregation required to identify the best fit for common purpose.

These real-world conditions provide no basis for expecting the retention guarantee absent from the simplified problems to reappear. Rather, non-linear and conflicting preferences, coupled performances, complex feasible spaces, increasing numbers of orientations, and declining dominance discrimination further separate numerical ordering in performance space from preference-consistent choice in design–decision space.

Preference-Guided IMAP Design–Decision Optimisation

Preference-guided IMAP design–decision optimisation reverses the conventional Pareto-based sequence. Pareto-based selection first filters the achievable set using numerical dominance and only then applies preference information to the retained alternatives. It can therefore identify the highest-preference alternative within the Pareto-filtered set but cannot guarantee that this alternative is the preference optimum over the complete feasible and acceptable design–decision space.

IMAP avoids this Pareto-filtering limitation by modelling and aggregating actor preferences before optimisation and using the resulting aggregated-preference function as the optimisation criterion from the outset. This requires no more preference information than a complete a posteriori procedure, which ultimately also requires such information to select from the Pareto set. The decisive difference is that IMAP applies preference information before alternatives are discarded, rather than afterwards, when only the retained alternatives remain available for selection.

IMAP represents actor preferences through performance-to-preference functions, aggregates them in a common preference domain using the A-FINE aggregator, and maximises the resulting aggregated preference over the complete feasible and acceptable design–decision space [15, 16, 17, 13]. Through the composite mapping from design–decision variables to performances and from performances to preferences, IMAP accommodates coupled performances, multiple actors, non-linear and non-monotonic preference functions, and conflicting actor preferences concerning the same performance.

IMAP is therefore recommended for preference-consistent design–decision optimisation. Numerical optimisation, approximation, or surrogate modelling may still support the search, but aggregated actor preference should define the optimisation criterion from the outset. Pareto filtering should not precede this preference-guided search, because it cannot guarantee retention of the preference optimum.

Conclusions

Pareto’s preference-independence claim that alternatives can be ordered independently of preference information has already been shown to be incorrect. This study establishes the stronger result that subsequently applying the same actor-preference information a posteriori does not restore a general guarantee of decision validity, because the preference optimum may already have been excluded by the preceding numerical dominance filter. This remains the case even when the union of the Pareto fronts under all possible minimisation–maximisation orientations is considered.

As a preference-independent method, Pareto analysis is therefore limited to numerical ordering and exploration in performance space. Pareto non-dominance does not imply preference optimality, and preference-consistent selection from a Pareto set does not guarantee identification of the best fit for common purpose. Consequently, Pareto-based selection cannot, in general, provide trustworthy, decision-valid multi-criteria decision support.

Preference-consistent design–decision optimisation requires actor preferences to define the optimisation criterion from the outset. IMAP implements this principle by modelling and aggregating actor preferences a priori and maximising the resulting aggregated preference over the complete feasible and acceptable design–decision space. The overarching conclusion is therefore unequivocal: numerical Pareto ordering cannot replace preference-consistent deciding, nor can its design–decision deficiency be corrected by introducing preference information only after Pareto filtering. As a final caution, decision-makers should not entrust conventional Pareto-based approaches, including contemporary LLM-supported Pareto implementations, with identifying the best fit for common purpose. Numerical non-dominance is not a trustworthy substitute for preference-consistent choice and may exclude the best fit for common purpose, potentially leading to suboptimal design decisions.

Data Availability

The current Preferendus software environment includes Preferendus-IMAP-IGS, which uses the A-FINE aggregator’s decision-valid preference-aggregation method to support preference-consistent design–decision optimisation. The software is publicly available through the Preferendus GitHub organisation: https://github.com/Preferendus.

The implementations of Problems 1–6 presented in this paper are available from the following GitHub repository: https://github.com/Preferendus/IMAP-vs-Pareto.

References

  1. [1]Ahmadianshalchi, A., Belakaria, S., & Doppa, J.R. (2023). Preference-aware constrained multi-objective Bayesian optimization. arXiv preprint. https://doi.org/10.48550/arXiv.2303.13034
  2. [2]Barzilai, J. (2010). Preference Function Modelling: The Mathematical Foundations of Decision Theory. In M. Erh gott, J. R. Figueira, & S. Greco (Eds.), Trends in Multiple Criteria Decision Analysis. Springer. https://doi.org/10.1007/978-1-4419-5904-1_3
  3. [3]Barzilai, J. (2022). Pure Economics. ISBN 9781039142121.
  4. [4]Deb, K., Pratap, A., Agarwal, S., & Meyarivan, T. (2002). A fast and elitist multiobjective genetic algorithm: NSGA-II. IEEE Transactions on Evolutionary Computation, 6(2), 182–197. https://doi.org/10.1109/4235.996017
  5. [5]Deb, K., & Jain, H. (2014). An evolutionary many-objective optimization algorithm using reference-point-based nondominated sorting approach, Part I: Solving problems with box constraints. IEEE Transactions on Evolutionary Computation, 18(4), 577–601. https://doi.org/10.1109/TEVC.2013.2281535
  6. [6]Ehrgott, M. (2005). Multicriteria Optimization (2nd ed.). Springer, Berlin, Heidelberg. https://doi.org/10.1007/3-540-27659-9
  7. [7]French, S. (2023). Reflections on 50 years of MCDM: Issues and future research needs. EURO Journal on Decision Processes, 11, 100030. https://doi.org/10.1016/j.ejdp.2023.100030
  8. [8]Marler, R. T., & Arora, J. S. (2004). Survey of multi-objective optimization methods for engineering. Structural and Multidisciplinary Optimization, 26, 369–395. https://doi.org/10.1007/s00158-003-0368-6
  9. [9]Martins, J.R.R.A. and Ning, A. (2022). Engineering Design Optimization. 2nd ed. Cambridge University Press, Cambridge, UK.DOI
  10. [10]Miettinen, K. (1999). Nonlinear Multiobjective Optimization. Kluwer Academic Publishers, Boston, MA, USA. https://doi.org/10.1007/978-1-4615-5563-6
  11. [11]Purshouse, R.C., & Fleming, P.J. (2007). On the evolutionary optimization of many conflicting objectives. IEEE Transactions on Evolutionary Computation, 11(6), 770–784. https://doi.org/10.1109/TEVC.2007.894202
  12. [13]Teuber, L.G., van Heukelum, H.J., & Wolfert, A.R.M. (2025). Advancing strategic planning and dynamic control of complex projects. arXiv. https://doi.org/10.48550/arXiv.2408.12422
  13. [14]Teuber, L.G., & Wolfert, A.R.M. (2024). Confronting Conflicts to Yes: Untangling Wicked Problems with Open Design Systems. arXiv. https://doi.org/10.48550/arXiv.2409.10549
  14. [15]Timp, L.J.K., Yorke-Smith, N., & Wolfert, A.R.M. (2026). Solving Highly Constrained Multi-Objective Decision Problems with the IMAP-IGS Preference-Guided Metaheuristic Framework. arXiv. https://doi.org/10.48550/arXiv.2607.17304
  15. [16]Van Heukelum, H. J., Binnekamp, R., & Wolfert, A.R.M. (2024). Socio-technical systems integration and design: a multi-objective optimisation method based on integrative preference maximisation. Structure and Infrastructure Engineering, 22(1), 36–53. https://doi.org/10.1080/15732479.2023.2297891
  16. [17]Wolfert, A.R.M. (2026a). Preference-based optimisation for integrated design and group decision-making. arXiv. https://doi.org/10.48550/arXiv.2603.19050
  17. [18]Wolfert, A.R.M. (2026b). Unique preference aggregation for multi-criteria design and decision-making. arXiv. https://doi.org/10.48550/arXiv.2601.19759

Paper details

vlnc:W_s5abqv5j
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 7

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents