Couple Simulation: A Novel Approach for Evaluating Models of Human Mate Choice

Traditional Social Psychological Approaches to Mate Choice

Although most relationships research within social psychology focuses on ongoing romantic relationships, there is a long history of social psychological theorizing on the determinants and processes of romantic partner choice. Three variables loom large in traditional social psychological theories of mate choice. The first is proximity: People are likely to enter into romantic relationships with those who live near them and with whom they frequently interact (e.g., Bossard, 1932). This occurs presumably both in part due to access and because familiarity increases attraction (Zajonc, 1968). The second variable is similarity: People tend to pair with self-similar others, at least along dimensions such as religious or political attitudes (Newcomb & Svehla, 1937). Attraction has long been observed to grow with the proportion of self-similar characteristics (Byrne & Nelson, 1965). Dissimilarity produces romantic repulsion as well (Rosenbaum, 1986). The third important variable is reciprocity in liking: In interpersonal relationships in general, and in romantic relationships in particular, we tend to like those who like us (e.g., Kenny & La Voie, 1982)—especially potential partners who like us more than they like other people (Walster et al., 1973). Furthermore, these three variables are not necessarily independent. For instance, preference for similarity might emerge in part because self-similar others are anticipated to reciprocally like us more (Condon & Crano, 1988).

Interdependence theory provides a broader theoretical framework from which much social psychological theorizing on mate choice derives (Thibault & Kelley, 1959). Interdependence theory conceptualizes relationship seeking and maintenance behavior in terms of people managing the rewards they receive and costs they endure as part of their romantic relationships. According to interdependence theory, a person compares the net rewards that they receive or that they expect to receive with two values: their comparison level (CL), or the net rewards a person believes they deserve from their relationship, and their CL for alternatives (CL_alt), roughly the net rewards expected from a person’s next best outside option. These comparisons are thought to regulate people’s satisfaction with their relationships and their motivation to abandon their relationships, respectively.

Interdependence theory’s emphasis on maximizing rewards and minimizing costs has particular relevance for the processes of partner selection. If all people attempt to maximize their net benefits by only staying in the most beneficial of their available options, then people who offer the greatest number of rewards at lowest cost to others will tend to be the most sought-after relationship partners. These people will have the greatest number of options to choose among, and so can choose the partners that offer them the greatest net rewards in return. To the degree that there is agreement about who provides the greatest rewards, the result will be assortative mating: Partners will tend to have correlated desirabilities such that the most consensually favorable people tend to pair with one another (Ellis & Kelley, 1999; Rusbult, 2003). This assortative mating importantly offers an alternative explanation for similarity between romantic partners to preference for similarity per se: People may be similar to their partners on desired dimensions such as physical attractiveness because desirable people tend to choose desirable partners (Kalick & Hamilton, 1986).

Interdependence theory also places special emphasis on mutuality and interactions at the dyad level, rather than at the person level. Central to interdependence theory is the fact that the rewards and costs a person receives from their relationship depends not only on their own behavior but on the decisions and behavior of their partner as well. Understanding relationships, partner choice included, therefore requires attention to the whole relationship, person and partner. The mutuality at the core of interdependence theory also creates the need for risk regulation (Murray et al., 2006). Reaping the rewards of a relationship requires seeking and building a connection with a romantic partner; however, doing so also poses the risk of costly rejection by that partner. Risk regulation theory proposes that people calibrate their relationship promotion behaviors to their perceived risk of rejection, seeking closeness when acceptance appears likely, but seeking self-protection when rejection seems likely. Risk regulation can, among other things, offer an explanation for the importance of reciprocity in partner choice: Two partners’ risk regulation systems can become calibrated to one another, with one partner’s relationship promoting behavior providing the other partner assurance that rejection is unlikely, promoting prorelationship behavior in return. Reciprocal exchange of these assurances will deepen as interdependence grows and the costs of rejection grow with it. Through such attention to rewards and benefits at the dyad level, interdependence theory offers explanations for important phenomena at the relationship initiation stage and beyond.

Evolutionary Approaches to Mate Choice

Evolutionary approaches, spearheaded by evolutionary psychologists and anthropologists, draw on natural and sexual selection as guides for theory building on the nature of human mate choice. Central to these approaches is an evolutionary psychological model of mind, according to which the brain is a natural computer composed of many specialized information processing systems (Cosmides & Tooby, 1995). These information processing systems are themselves expected to be adaptations designed by selection to solve “adaptive problems”: historically recurrent problems faced by human ancestors that, if solved, would have increased reproduction. Each of the mind’s psychological adaptations is expected to capture information from its environment, process that information, and produce behavioral, affective, or cognitive outputs that would have historically been tributary to solving some recurrent problem faced by human ancestors (Cosmides & Tooby, 1995). This form-function link lends evolutionary theory its heuristic value to psychologists: If one has an understanding of the adaptive problems faced throughout human evolutionary history (function), one also has clues to the likely information processing adaptations contained within the mind (form).

A key theory for evolutionary perspectives on mate choice is Trivers’s parental investment theory (Trivers, 1972). At the core of parental investment theory is an acknowledgment that producing and rearing offspring requires the expenditure of finite resources, chiefly time and energy, rendering reproduction inherently costly. However, the costs of reproduction are not equal for all individuals. The variance in reproductive costs is meaningful because, all else equal, greater reproductive costs means scarcer lifetime reproductive opportunities. Those individuals for whom reproduction is costlier face a problem of reaping maximum benefits from few opportunities and therefore benefit from more judicious mating than those for whom reproductive opportunities are plentiful. In most species, females systematically experience greater reproductive costs, explaining the general pattern of sexual dimorphism across species in which males are typically larger in size, exhibit more extravagant ornaments and display behaviors, and mate more indiscriminately, whereas females are typically smaller in size, cryptic, and more selective in mate choice. The notable exceptions to this general trend are “sex-role reversed” species, such as phalaropes, in which males bear greater reproductive costs, and consequently—and in accordance with parental investment theory—tend to be smaller and drabber than the flashier females (Trivers, 1972).

Of relevance to parental investment theory, human reproduction is particularly costly (Aiello & Key, 2002; Valeggia & Ellison, 2009), owing in part to the heavy energetic demands of human infant brains (Ellison, 2009; Kuzawa et al., 2014) and the difficult nature of human foraging (Kaplan et al., 2000). This is thought to be one reason why our species is one of the very few mammal species to engage in long-term, pair-bonded mating in which males and females cooperate to rear offspring, jointly shouldering our high costs of reproduction (Kaplan et al., 2001; Key & Aiello, 2000). On the basis of parental investment theory, this unique long-term mating strategy is predicted to have favored the evolution of selectivity in romantic mate choice in both males and females (Buss & Schmitt, 1993; Trivers, 1972). Inspired by this prediction, the predominant emphasis within the evolutionary literature on long-term mate selection has been on documenting the nature of this selectivity, especially as guided by the meta-hypothesis that the content of modern human desires will reflect features that would have been beneficial to human ancestors (Buss, 1989; Symons, 1979).

Furthermore, parental investment theory itself can be conceptualized as a special case of a broader theoretical framework within evolutionary theory known as life history theory (Kaplan et al., 2001; Stearns, 1992). Although primarily emphasized by evolutionary anthropologists, life history theory has been taken up by psychologists as well—albeit not without some controversy (Del Giudice, 2020; Nettle & Frankenhuis, 2019; Zietsch & Sidari, 2019). Life history theory begins with the observation that all organisms fundamentally face a problem of capturing energy from their environments and applying that energy to reproduce successfully. But solving this problem requires first accomplishing a variety of subtasks: growing a body to maturity, defending that body from pathogenic infection, repairing that body from environmental damage, foraging successfully for resources, gestating offspring, parenting offspring, and so on. The energy available to accomplish all these vital tasks is finite, and resources allocated to one task cannot be allocated to another. This forces trade-offs, for instance, between future and current reproduction: Energy spent on growing one’s body may facilitate greater reproduction in the future, but comes at a cost of energy that could have been spent reproducing now. Organisms must evolve adaptations that prudently manage allocation of finite resources to these competing tasks based on the expected fitness returns of alternative investment portfolios (Kaplan et al., 2001).

One classic life history trade-off is between the quantity of offspring produced and the quality of investment offered to each. Human offspring in particular require substantial parenting to be able to successfully reproduce themselves (Kaplan et al., 2000); however, all energy invested in parenting one offspring is energy that could have been spent producing new offspring. Mating and parenting consequently trade off (Hurtado & Hill, 1992; Marlowe, 1999; Valeggia & Ellison, 2009), where time and energy spent searching for new mates appears to come at a cost to parenting and vice versa. Life history theory therefore connects mate selection intrinsically to all the problems of life; to a life history theorist, understanding mate selection requires attention not to just mating itself, but to the whole-organism balance of life history trade-offs.

Cognitive Science Approaches to Mate Selection

A final theoretical approach worthy of brief mention is what I will call the “cognitive science” approach. These are approaches that draw on theories and methods from disciplines such as judgment and decision-making, economics, statistics, and computer science to attempt to uncover the decision processes involved in human mate selection. Generally, the most formal of the theoretical approaches to mate choice, cognitive science theorists less frequently study real-life mating behavior, but rather engineer simplified scenarios that capture essential elements of real mate choice while abstracting away details. These scenarios can then serve as conceptual laboratories for developing and testing decision-making strategies.

The cognitive science approach does not offer so much as an alternative theoretical perspective to the prior approaches; rather, it is complementary. In a sense, whereas social psychological and evolutionary approaches more heavily emphasize the proximate and ultimate tasks human mating psychology attempts to solve, cognitive science approaches focus on offering precise models of the information processing systems the mind could use to accomplish those tasks. For this reason, cognitive science approaches are often integrated with social psychological and/or evolutionary approaches.

One classic and well-studied example of such a scenario is the “secretary problem” (also known as the “dowry problem”; see Todd & Miller, 1999). This scenario imagines a hiring manager (or a bachelor/ette) attempting to hire the best secretary (or choose the best mate) out of a pool of eligible candidates. Candidates can be evaluated only one at a time, simulating, for instance, a sequence of dates with different potential partners. Each candidate has an overall quality, visible only after evaluation, and which may be a function of a variety of characteristics (e.g., computer skills and time management or kindness and physical attractiveness). Once a candidate is selected from the pool, the decision is final—reflecting the relative permanence of romantic bonds; furthermore, once a candidate is declined, they cannot be returned to later—approximating the scorn of a spurned suitor.

Decision strategies can be tested on this basic problem structure and evaluated on the basis of the quality of their chosen candidates. For example, satisficing strategies perform well in such sequential search problems. These strategies search through an initial subset of candidates to observe the distribution of candidate quality. Afterwards, they select the first candidate who is as good or better as the best candidate observed in the initial search (Miller & Todd, 1998). Although developed on artificial scenarios, decision strategies developed from cognitive science approaches can be compared with decisions made by real humans in analogous scenarios (e.g., see Beckage et al., 2009; Brandner et al., 2020).

A critical distinction among cognitive science approaches is between optimizing and satisficing strategies (Gigerenzer & Todd, 1999). Optimizing strategies follow more directly from the traditional economic, utility maximizing perspective. These strategies are designed to make decisions so as to optimize some important criterion: for instance, maximizing rewards of a given partner choice, minimizing search costs, maximizing the stability of mated pairs, or combinations thereof. Satisficing approaches instead follow more from a perspective of bounded rationality. Satisficing strategies do not attempt to produce choices that are necessarily optimal on any dimension; rather, they are designed to produce choices that are “good enough.” The motivation for such strategies is the recognition that human computational power is finite: We have only so much attention, so much memory, so much time, and many demands on these limited computational resources. Satisficing strategies can require minimal use of finite resources while still producing performance comparable to more complex and demanding optimizing strategies, rendering them potentially more feasible and more biologically plausible alternatives to optimizing strategies (Miller & Todd, 1998).

The savings in complexity afforded by satisficing strategies are argued to more than make up for any deviations from optimality. For example, within the secretary problem, the optimal strategy for maximizing one’s probability of selecting the very best mate available is to search through the first 37% of one’s expected mate pool without making a choice and then afterward picking the first mate who is as good or better than the best mate encountered in the initial search. However, a satisficing strategy of searching for just a good mate—but not necessarily an optimal mate—can find higher quality mates on average using shorter search times (Miller & Todd, 1998). For this reason, many theorists in the cognitive science tradition, especially those in the satisficing tradition, place a heavy emphasis on constraints on choice and on simplicity of decision-making strategies.

Computational Models Can Help Solve the Theoretical Challenges of Mate Choice

Computational models can facilitate progress toward answering this question by solving many of the problems faced in theorizing about mate choice. Computational models do not constitute distinct theories from verbal models nor do they constitute empirical data in and of themselves. Rather, computational models represent a mode of theorizing. In a computational modeling framework, rather than imagining a counterfactual world and deriving predictions through reasoning, the researcher instead expresses their theoretical assumptions in computer code—specifically, code that literally creates a simulated world in which the researcher’s theoretical assumptions are true. The researcher can then derive predictions by directly observing and measuring features of this simulated world. Theoretical assumptions can still spring from any of their traditional sources: for instance, intuition, observation, or from meta-theories such as evolutionary theory. And just as with verbal models, model predictions can be compared against the real world to determine their veracity and thereby assess the plausibility of the model’s underlying assumptions.

How do such models help solve the problems faced by theories of mate choice? There are at least three key benefits to expressing theory in terms of computational models: computational models help researchers generate predictions, they facilitate communication between theorists, and they more closely represent the operation of the mind.

The Complexity of Mate Choice

Consider building a robot that can select mates from a human mating market. On its path to love, this robot will encounter several problems (Figure 1). First, the robot will encounter an array of potential romantic partners who vary across an infinite number of dimensions. Some potential partners will have more symmetrical faces, others will have more symmetrical arm hair; some will be wealthier in cash, others will possess larger stamp collections; and some will be healthier, others will have healthier cousins. There is no end to the number of dimensions along which the robot could compare its potential mates, but the robot must make a choice in finite time.

OPEN IN VIEWER

Figure 1.

A diagram of the mate choice process, highlighting the three key tasks of forming preferences, integrating preferences, and navigating dynamic mating markets. Mate preference integration and mate choice decision algorithms are represented as gray and black boxes, respectively, to reflect the current state of knowledge on these processes.

The choice of which dimensions deserve attention could, to the robot, be arbitrary. As far as the robot is concerned, why should it care more or less about a potential mate’s financial wealth than their philatelic wealth? However, to the engineer of real-world mate choice psychologies—natural selection—the choice of comparisons is far from trivial. In the eyes of selection, the true determinant of a potential mate’s “value” as a romantic partner is the reproductive benefits that mate would offer if they were successfully attracted (Sugiyama, 2015). Individuals whose value judgments correlated with these reproductive benefits would select mates who allow them to produce more offspring on average, increasing the representation of whatever genetic endowments contributed to the development of their judgment mechanisms. As a result, selection’s sieve would swiftly replace psychologies ignorant of this mate value dimension with psychologies that are not. If our robot’s mate choice psychology is to emulate real human mate choice psychology, we must design it so that its value judgments are attuned to mate value. And yet, only an infinitesimally small subset of all perceivable differences would be useful to the robot as indicators of mate value.

To successfully select a mate out of this sea of variation, the robot must first establish preferences. That is, the robot needs some ability to form abstract trait-relevant representations that enable it to (a) identify, out of the infinite possibilities, which trait dimensions are relevant to evaluate and (b) assign valence to each of these dimensions, such that some regions of the trait space are evaluated favorably relative to other regions. The “preferences” themselves—the psychological parameters that dictate evaluations as a function of perceptions—could be implemented in a variety of ways, including ideal point values, preferred ranges, or as preference functions that describe how trait values should be transformed into valenced evaluations.

The robot next needs a means to apply its preferences to estimate its potential mates’ mate values. Here, the robot faces a second challenge: Potential mates vary at least somewhat independently across preference dimensions. This means that a mate evaluated positively on one preference dimension is not guaranteed to evaluate as positively on another. A corollary of this is that any given potential mate will be preferable on some dimensions but not on others. To select among its set of imperfect potential options, the mate choice robot needs to somehow integrate its individual evaluations across dimensions to form an overall assessment of each potential mate under consideration. The robot could use all its individual evaluations at this stage or relatively few. This evaluation process could also take several forms, including checklists, weighted sums, similarity functions, or other decision heuristics. And the output evaluations could exist in a variety of formats, from a continuous value estimate, to an ordinal ranking, to a binary “go/no go” signal.

Finally, once the mate choice robot has evaluated one or more potential mates, it needs to apply these evaluations to determine which mates to pursue and ultimately select. Here, in addition to no option being perfectly appealing, the mate choice robot faces two new problems. First, no one is ever the only person searching on the mating market; the mate choice robot will now experience competition from mating rivals. To the degree that the robot’s preferences overlap with its competitors, it may find its most preferred partners removed from the market by rivals that are either more appealing, more decisive, or both. Relatedly, the robot will face a second problem in that human mate choice is mutual. It is not enough for the robot to determine the partner it prefers the most; that partner must also pick the robot in return. And, of course, there is no guarantee that the robot’s most preferred partners will return the robot’s affection.

This third task of navigating the mating market poses the most complex of all the problems the mate choice robot will encounter as the optimal solutions depend on the distribution of potential mates, the distribution of rivals, and the robot’s own traits and preferences. The robot must also take care to search the mating market efficiently. The robot has only finite time and energy to dedicate to mate evaluation and pursuit; in addition, the robot must manage opportunity costs as the resources dedicated to pursuing one mate require forgoing potentially superior options. The robot must balance exploiting known options and exploring less-known options to find suitable partners while minimizing search costs. Further complicating these issues, on the mating market, the robot is embedded within a dynamic social system, where any individual’s decisions affect the decisions available to everyone else. Navigating this stage of mate choice requires decision rules that govern evaluations, mate pursuit, and/or mate rejection contingent on the robot’s shifting circumstances.

Solving the Problems of Mate Choice

These are some of the key problems people face in the course of selecting a romantic partner. But how, precisely, does human mating psychology solve these problems? The answer to this question has been of interest to human mating researchers for decades. However, research effort has not been distributed equally across the separate tasks of choice. Consequently, human mating research has made more progress in understanding some aspects of the design of human mating psychology than it has for others.

Actual mate choice

Whereas the psychology of mate preference integration is becoming faintly illuminated, the area of human mating that remains most in the dark is perhaps the most important: actual choice. The designs of human mate preference psychology and preference integration psychology reveal how people evaluate their potential mates. However, they do not tell us how people use these evaluations to select actual romantic partners out of their pool of potentials. This is the most challenging task of mate choice, as here individuals must navigate dynamic mating markets, constrained by the availability of partners, the behaviors and choices of romantic rivals, and the challenge of securing mutual attraction from preferred potential mates.

The complexity of this final task is likely responsible for the slow development of computational models of the decision processes involved in actual mate choice. Any model that can accurately describe how people navigate their mating markets will necessarily propose several interacting processes that are difficult to measure, manipulate, or observe. To illustrate, here I will consider four exemplar models of mate choice: (a) the Kalick-Hamilton model, (b) the aspiration threshold model, (c) the Gale-Shapley algorithm, and (d) the resource allocation model. This is by no means an exhaustive list. However, the first three of these models are the most established models in the literature; most other models are variations on their major themes. The RAM is a new model proposed here that fills important gaps left open by the first three.

The Kalick-Hamilton Model

The first model I will consider, the Kalick-Hamilton model (KHM), derives from social psychological theory on mate choice. This model was first proposed in Kalick and Hamilton’s (1986) classic paper on assortative mating. A robust finding from the interpersonal attraction literature is that people tend to mate somewhat assortatively for physical attractiveness: Attractive people tend to have attractive partners. Kalick and Hamilton (1986) sought to provide an explanation for this observation. At the time, the dominant hypothesis for assortative mating was a preference for similarity: that people desire partners who are close in attractiveness to themselves. Kalick and Hamilton tested an alternative explanation: Mating markets are competitive, and the people who are most in demand (i.e., physically attractive people) are most likely to successfully pair with partners they most desire (i.e., physically attractive partners). To evaluate the plausibility of this hypothesis, they set up a model of a mating market incorporating competitive, mutual, and probabilistic mate choice. This model represents a computational expression of the long-standing prediction of assortative mating from interdependence theory (Ellis & Kelley, 1999; Rusbult, 2003).

Overall, their model is simple. At each time step, agents pair into dates at random. On these dates, each agent first assesses the attractiveness of their paired partner. Next, agents decide whether to make an offer of commitment to their date partners based on the result of a weighted coin flip. The probability of each agent making an offer of commitment is proportional to two things: the attractiveness of the agent’s partner and the number of dates the agent has been on already. As the agent’s partner becomes more attractive, a decision to commit becomes more likely; furthermore, as the agent goes on an increasing number of unsuccessful dates, it becomes less discriminating and increasingly likely to offer commitment to less attractive partners. If two agents make mutual offers of commitment, they pair and exit the mating market. If either declines, the agents end their date and reenter the mating market. This pairing and coin flipping process iterates until all possible agent couples form. Although this model was designed to explain assortative mating for attractiveness, it can be extended to mate choice based on multiple preferences: If one has a model of mate preference integration, one can replace “physical attractiveness” in the decision process with overall mate value evaluations.

The KHM has several desirable features as a computational model of human mate choice. First, it is algorithmic: It provides a set of unambiguous steps that can move a population of single individuals into a population of romantic pairs based on their mate preferences. A population of mate choice robots that followed the Kalick-Hamilton choice algorithm could successfully solve all the problems real people encounter in selecting a romantic partner. Second, the KHM accomplishes this pairing while respecting the mutual nature of human mate choice: All individuals get to choose whom they pair with and each person’s choice matters in their outcome.

Nonetheless, this model is relatively simplistic. It is also highly stochastic, in terms of both initial mate approach and final mate choice. Agents are paired with random partners at each time step and thus are no more likely to encounter highly appealing potential mates than they are to encounter highly unappealing partners—a seemingly unrealistic assumption. Furthermore, the final mate choice comes down to a simple, albeit weighted, binomial. This means that a proverbial “10” always has some probability of committing to a “1” and a “1” always has some probability of declining an interested “10.” Stochasticity per se is not undesirable—some degree of stochasticity in mate choice might make functional sense for solving specific problems encountered in mate choice, and so some stochastic elements may be realistic and desirable features of mate choice models. Nonetheless, given the central importance of mate choice to reproduction, it seems highly unlikely that selection would craft such a mate choice psychology as random as the KHM implies.

The Aspiration Threshold Model

Moving up one step in complexity, we can consider what I will call the “aspiration threshold model” (ATM). This is an abstraction of a model proposed in several forms (e.g., Miller & Todd, 1998; Simão & Todd, 2003; Todd et al., 2005) but is closest in form to (but still somewhat different from) the version proposed in Todd and Miller (1999). The ATM follows from a merger of the evolutionary and cognitive science approaches to mate choice theory, and in particular falls within a satisficing approach to decision-making.

In this model, individuals are assumed to have personal aspiration levels: thresholds of mate value which potential mates must meet and which are learned through experience on the mating market. Agents begin their mating careers with uninformative thresholds—for instance, seeking at least a “5” on the proverbial 10-point mate value scale. Then, just as in the KHM, each agent goes on a date with a random partner at each time step. If an agent is single, it will make an “offer” to any potential mate whose mate value exceeds their aspiration threshold. Agents temporarily pair with their date partners if they make mutual offers of commitment. After agents enter a temporary pairing, they continue to go on dates with new random partners, but adjust their decision rules slightly: Now they will make an offer to a new partner only if that new partner both exceeds their aspiration threshold and is higher in mate value than their current partner. If this new offer is accepted, the agent breaks up with their current temporary partner and enters a temporary relationship with the new mate. Temporary pairs become permanent once agents have stayed paired to one another for a preset duration.

At the same time that individuals are making pairing decisions, they are also calibrating their aspiration thresholds. Each agent in this model pays attention to the offers they receive from others and adjusts their aspiration threshold accordingly. When an agent does not receive a commitment offer from a potential mate who is below their threshold, suggesting they have been too ambitious, they adjust their threshold downward slightly; when they receive commitment offers from mates above their thresholds, suggesting they have been too conservative, they adjust their thresholds upward. Iterating this learning process across random dates rapidly yields aspiration thresholds that are finely tuned to the agents’ mate values (see Supplemental Figure 1). These thresholds then help agents pair into good matches, given their mate values.

Like the KHM, the ATM is appealing as a model of mate choice in that it is algorithmic and incorporates mutual mate choice. Furthermore, the proposed learning process is elegantly efficient. This model is still relatively stochastic; however, it has two features that render it less stochastic than the KHM. First, whereas the ATM is still stochastic in terms of mate approach—individuals pair into dates randomly, and no structures in the environment or the agents’ psychology bias these random encounters toward high or similar mate value partners—it is not stochastic in terms of actual mate choice: In the ATM, choice is governed by a deterministic aspiration threshold. Second, the ATM allows for mate switching: Agents can “hold on” to an appealing partner provisionally but then “trade up” if a better, mutually interested partner appears. This mate switching process might help to smooth out some of the randomness introduced by random mate approach, yielding an algorithm less likely to choose relatively low mate value partners or deny relatively high mate value partners.

The Gale-Shapley Algorithm

In contrast to the KHM and ATM, the Gale-Shapley Algorithm (GSA) offers a fully deterministic model of human mate choice. Fitting most with the cognitive science approach to mate choice, and in particular a traditional, rational utility maximizing approach, this algorithm is an economic model designed to solve the “stable marriage problem” (Gale & Shapley, 1962). Despite the name, this is not a problem of mate choice per se but rather is a general problem of sorting lists into pairs according to ordered preferences: for instance, pairing medical school students who vary in qualifications with residency positions that vary in prestige. The algorithm proceeds in a series of steps. First, each unpaired male makes an offer to his most preferred female. If an unpaired female receives any offers, she temporarily pairs with whichever suitor she prefers most and rejects any others. If a paired female receives an offer, she compares her current mate with her suitors. If she prefers her current partner to all her suitors, she rejects each suitor and stays with her current mate; if she prefers a suitor to her current partner, she jettisons her mate and pairs with the new male, rejecting her former partner and all other suitors at the same time. Any unpaired males next make new offers to their most preferred females among those who have not yet rejected them. This process repeats until all possible couples form.

Like the prior two models, the GSA is algorithmic and unambiguous. Unlike the prior models, it is completely deterministic: It is nonrandom in both mate approach and mate choice and, given the same start conditions, will yield the same pairs each time it runs its course. The GSA also has a potential advantage in that it is guaranteed to find pairs that are “stable”: Among all pairs, no individual will prefer to be with another partner who would also prefer them in return. This stability may or may not be a feature of real human mate choices, but could be useful for applications of mate choice models in, for instance, making partner recommendations.

Nonetheless, this algorithm suffers a critical limitation in that it is inherently asymmetric: Only one sex actively pursues their preferred partners. Females only act on their preferences when they have competing offers from males. As a result, pairings are more optimal with respect to male preferences than they are for female preferences. Such asymmetry in choice could be a reasonable assumption for a variety of nonhuman species, but is not likely consistent with human mate choice (Stewart-Williams & Thomas, 2013).

The Resource Allocation Model

Each of these models differs on a number of dimensions, for instance, stochasticity, symmetry, and overall complexity. However, they each share one noteworthy feature: They are all “binary” in the sense that they all assume that, at a given time point, each individual is either fully pursuing a given potential mate or is not pursuing them at all. This is a somewhat peculiar assumption, given that human attraction at least appears to be continuous: people are strongly attracted to some potential mates, moderately attracted to others, and weakly attracted to others still.

To address this, here I propose a model new to the literature, called the resource allocation model (RAM), designed to capture this continuous nature of attraction. It draws loose inspiration from life history theory. Life history trade-offs are most often conceptualized as trade-offs between investment in different fitness-relevant processes—for instance, between an organism investing its limited resources in growing its body larger, maintaining that body, pursuing mates, or parenting offspring. This perspective is illuminating. But we can also “zoom in” on the resources invested into each process and see that investment trade-offs have a hierarchical nature. With the time and energy invested in mate pursuit, we have a finite set of resources that can now be invested across a set of potential mates. These investments face a similar problem of trade-offs: every hour spent courting one potential mate is an hour that cannot be spent courting another. Therefore, just as life history trade-offs are conceptualized as a process of deciding how to allocate finite resources across a set of vital processes, mate choice can be thought of as a process of deciding how to allocate finite resources across a set of potential mates.

The RAM assumes that two things are relevant to this allocation decision: the mate value of each potential mate and their likelihood of reciprocally investing their finite resources in return. Allocation in this model occurs in two phases. First, each agent allocates their limited mating resources to their potential mates in direct proportion to their relative mate values. Mates evaluated as high in mate value receive a greater initial proportion of an agent’s resources (e.g., the suitor commits more of their time to higher mate value partners), whereas mates low in mate value receive fewer resources. At this stage, resources are spread wide and thin, and differential allocations could be taken to represent differences so small as taking the time to make a flirtatious comment to a particularly appealing partner.

After this initial allocation, however, agents reallocate their resources in proportion to their degree of mutual investment shared with each mate. Agents disinvest from mates who did not return their initial investment and shift these freed-up resources to mates who showed stronger mutual interest because no social partner is more valuable than a partner who values you in return (Tooby & Cosmides, 1996). Iterating this reallocation process causes agents to gradually drift from broadly investing resources in many potential mates to heavily investing in few, reciprocally interested partners (Figure 2). Compared with the small allocations given to potential mates at the start of the resource allocation process, the relatively large allocations given to specific partners by the end of mate choice could be taken to represent the substantial time and energy dedication demanded by a committed romantic relationship. This RAM therefore provides a model of human mate choice that is unique among the prior models in that it is simultaneously symmetrical, deterministic, and continuous.

OPEN IN VIEWER

Figure 2.

A social network representation of mate choice according to the resource allocation model. Nodes represent individual agents and edges represent the strength of mutual investment between agents, with edge width proportional to the product of investment given and received. Agents initially broadly invest in many potential mates (left), but as time progresses gradually concentrate their resources onto few, mutually interested partners (middle, right).

As a novel model, the RAM integrates concepts from across the traditional theoretical approaches to human mate choice. This demonstrates the value of computational models for building novel theories that integrate concepts from multiple perspectives. The RAM borrows from both evolutionary and cognitive science approaches an emphasis on navigating constraints on finite resources. It additionally unites these previously disparate senses of constraint by pointing out the nested nature of resource allocation: Life history trade-offs between competing processes are what constrain the resources available for mate choice decision-making. The RAM furthermore points out a symmetry between these two senses of trade-off: Life history trade-offs and on-the-ground mate choice both involve an analogous problem of deciding on an investment schedule that allocates finite resources across sets of differentially beneficial options so as to maximize returns.

The RAM additionally borrows from traditional social psychological theory by being the only model among the four considered here to explicitly emphasize reciprocity in the decision-making context. Directly analogous to risk regulation, agents in a RAM invest in a potential partner to the degree that they are confident that their partner will reciprocate their interest. This emphasis on reciprocity also allows the RAM to capture the emergence of romantic relationships over time in a way that binary choice models do not. Furthermore, compared with the three prior models, the RAM is the only model in which an agent’s own pursuit decisions are a joint function of both their own and their partner’s prior behavior, creating a dyadic emphasis not unlike interdependence theory’s. The incorporation of mate value into initial allocation decisions finally also helps to produce the assortative mating anticipated by interdependence theorists.

Evaluating and Comparing Models of Human Mate Choice

The four models above reflect computational expressions of theory from across the traditional theoretical approaches to understanding human mate choice. Table 1 summarizes these models in terms of their noteworthy features, theoretical origins, and theoretical emphases. Each of these models above could, in principle, approximate the set of decision rules used by human mating psychology to guide actual mate choice. But how, as mating researchers, can we decide which model is most plausible overall? This question has proven challenging to answer. Each model proposes several decision processes, some of which are challenging to manipulate or measure in the laboratory. For example, how would one reliably measure subtle changes in resource allocation across a person’s pool of potential mates, or the change in one’s aspiration threshold across a lifetime of mating experiences? Even worse for making clear comparisons, some models, such as the KHM, propose that final mate choice is at least partially stochastic, whereas others assume deterministic mate choices. Devising critical tests that can distinguish between any two of these models using real mate choice data is difficult enough. Finding the best model in the complete set is exponentially more difficult. Without tools that can facilitate this model evaluation, developing new, more accurate models of mate choice becomes challenging, and progress in the development of mate choice theory is slowed.

OPEN IN VIEWER

Table 1.

Origins, Features, and Key Theoretical Emphases of Mate Choice Models.

Model	Origin	Theoretical traditions	Probabilistic approach	Probabilistic choice	Continuous versus binary	Symmetrical	Noteworthy emphases
Kalick-Hamilton model	Kalick & Hamilton (1986)	Social psychological	Yes	Yes	Binary	Yes	Probabilistic choice, assortative mating, declining standards over time
Aspiration threshold model	Todd & Miller (1999)	Cognitive science and evolutionary	Yes	No	Binary	Yes	Learned aspiration threshold, satisficing
Gale-Shapley algorithm	Gale & Shapley (1962)	Cognitive science	No	No	Binary	No	Stable pairs, optimizing
Resource allocation model	Novel	Social psychological, cognitive science, and evolutionary	No	No	Continuous	Yes	Trade-offs, constraints, reciprocity, dyadic

The most common method for evaluating mate choice models within the literature is by assessing their ability to reproduce group-level statistics observed in human data. For instance, Kalick and Hamilton (1986) evaluated the plausibility of their model on the basis of its ability to reproduce the correlation in physical attractiveness commonly observed in human samples. Other studies have similarly evaluated different models of mate choice on the basis of their ability to reproduce correlations indicative of assortative mating (Conroy-Beam, 2018; Hitsch et al., 2010; Smaldino & Schank, 2012; Xie et al., 2015). Another common approach to model validation is to assess whether models reproduce broad demographic trends, such as rates of marriage as a function of age (French & Kus, 2008; Knittel et al., 2011; Smaldino & Schank, 2012; Todd et al., 2005).

These can be useful criteria. An accurate model of human mate choice must be able to yield, for instance, realistic marriage rates. Nonetheless, they are also relatively abstracted: Although mate choice models are models of individual choices, these comparison approaches evaluate models on their ability to reproduce observations at the population level. There is no guarantee that models that perform well in predicting aggregate trends also perform well in predicting the specific individual choices that make up the aggregate. Different models, with radically different underlying mechanics, can produce degrees of assortative mating or marriage rates that are broadly similar to one another and to human data even if they make quite different choices at the individual level. Equifinality is a challenge for any method of model comparison, but criteria as simple as assortative mating correlations render fine discrimination among different models of mate choice especially difficult. Despite this, model comparisons that evaluate the ability of alternative models to predict individual choices are relatively rare (although see Beckage et al., 2009; Brandner et al., 2020).

The absence of precise tools for evaluating models of mate choice, in conjunction with the complexity of mate choice itself, likely accounts for the relatively slow progress in this research area. By comparison, mate preference researchers have ample means to evaluate novel hypotheses concerning mate preferences, including survey methods (e.g., Fletcher et al., 1999; Li et al., 2002), archival data analysis (e.g., de Sousa Campos et al., 2002; Kenrick & Keefe, 1992), or revealed preference designs (e.g., Mogilski et al., 2014; Wood & Brumbaugh, 2009). Preference research has accordingly been fertile, with regular documentation of previously undiscovered mate preferences. However, a researcher who goes through the considerable effort of generating a novel mate choice model has no clear means to demonstrate that their new model is more or less plausible than any of the prior possibilities. Absent such a useful model comparison metric, human mating researchers cannot iteratively refine theoretical models of mate choice and build toward a more accurate understanding of human mating over time.

Model Setup

Model Results

Evaluating deterministic models with couple simulation

The ability of the couple simulation method to identify a population’s true mate choice algorithm is most straightforward for those populations in which mate choice was deterministic. These choice algorithms contain no stochastic elements; as such, given the same initial sample of individuals, these algorithms produce the same couples each time they run. Deterministic algorithms therefore make a singular prediction for each sample of couples and have a singular accuracy value that can be compared with the accuracy values of other models.

Figure 5 presents the results of the proof-of-concept models, focusing on the two deterministic mate choice algorithms: the GSA and the RAM. Three key findings are observable from this figure. First and most importantly, the couple simulation method does in fact correctly recover the true model of mate choice within each of these populations. When the GSA was the ground truth mate choice algorithm, this algorithm had significantly higher simulation accuracy across 100 random samples relative to the RAM at all sample sizes above k = 100, as indicated by the 95% confidence intervals (CIs). When the RAM was the ground truth mate choice algorithm, this model significantly outperformed the GSA in couple simulation at all sample sizes. It is also notable that the RAM appears more reproducible overall than the GSA: When the RAM is the true model, simulation accuracies are generally higher for both models and the RAM outcompetes the GSA at all sample sizes. This may occur because the RAM more heavily weighs all individuals’ preferences in making mate choices, unlike the asymmetric influence of preferences found in the GSA.

OPEN IN VIEWER

Figure 5.

Couple simulation accuracy for deterministic models of mate choice. (A) GSA population and (B) RAM population. Error bars represent 95% confidence intervals.

Second, although couple simulation performs well at reproducing sampled couples in general, simulation accuracy declines overall as sample size increases. In both populations, simulation accuracy begins relatively high at the smallest sample sizes, with both models reproducing more than 50% of the sampled couples when the sample size is k = 50 dyads. However, this performance declines in both populations, with each model reproducing less than 30% of couples when the sample size is k = 500. This decline in simulation accuracy was not predicted in advance and the reasons for its existence are not obvious. However, one probable contributor is that larger mating markets give agents more and potentially more appealing alternative options: With a larger sample size, the probability increases that the mating market contains a genuinely better match for each agent than their chosen partner. These superior matches represent partners the agent would have chosen had they been available when the agents made their original mate choices. Although these matches perhaps better reflect the agents’ desires, they also reduce model simulation accuracy by causing agents to deviate from their original mate choices.

Nonetheless, the third key observation available in Figure 5 is that although overall simulation accuracy tends to decline with sample size, discrimination between the true and the alternative model tends to increase. This is clearer for the GSA: In the two smallest sample sizes, the GSA fails to significantly outcompete the RAM even when it is the population’s true mate choice algorithm. However, the GSA has consistently greater simulation accuracy than the RAM once sample sizes increase beyond k = 100. This gain in discrimination is particularly clear when looking at the proportion of samples in which the true model was the best performing model (Table 2). When the sample size is just k = 50, the GSA outperforms the RAM in couple simulation in just 51% of samples, no better than chance. By the time the GSA is significantly outperforming the RAM on average, it is still the best performing model in just 68.5% of samples. However, once the sample size is at least k = 300, the GSA outcompetes the RAM in at least 80% of samples. Stated another way, a decision heuristic that assumes that the best performing algorithm is the true algorithm will yield the correct inference in at least 80% of randomly drawn samples of k = 300 dyads in which the true model is tested. These results indicate that if decision processes guiding mate choice are in fact deterministic, the couple simulation method can be used to identify relatively plausible models of mate choice and can go so far as to correctly identify a population’s ground truth mate choice algorithm, given just random samples of couples from the population.

OPEN IN VIEWER

Table 2.

Proportion of Samples in Which True Mate Choice Algorithm Had Highest Simulation Accuracy.

True model	Sample size (k dyads)
	50	75	100	200	300	400	500
Gale-Shapley algorithm	0.51	0.575	0.685	0.75	0.81	0.925	0.935
Resource allocation model	0.99	1	1	1	1	1	1

Evaluating stochastic models with couple simulation

Incorporating stochastic models of mate choice makes evaluating the validity of couple simulation more complex. Unlike deterministic algorithms, stochastic mate choice algorithms produce slightly different couple configurations each time they run even on the same initial sample of couples. For this reason, these stochastic models do not have a single couple simulation accuracy value; rather, they have a distribution of simulation accuracies. To compare the performance of stochastic models with other models, these accuracy distributions must be summarized by some value that allows for unambiguous but fair comparison of their overall performance. Initial intuitions for such a summary value might include the mean performance of a stochastic model or a model’s best performance across some number of model runs; however, both of these summary methods cause stochastic models to systematically underperform in couple simulation (see Supplemental Materials Section S2). A better summary method equates a stochastic model’s simulation accuracy with what I will call the model’s “best likely” performance.

The intuition behind the best likely performance summary is this: The observed simulation accuracy for a given run of a stochastic model of mate choice can be thought of as a random draw from a binomial distribution, where n is the number of couples in the sample and p is the unknown “true” accuracy of the model in the population. The best likely accuracy approach estimates the best likely value of p by using the model’s observed average performance to parameterize a beta distribution from which the p estimate is selected. To apply this approach in the proof-of-concept models, I ran each stochastic model 100 times on each sample of agent couples and calculated the average percentage of couples accurately reproduced by each stochastic model across 100 model runs. I then parameterized a beta distribution with α equal to this average percentage and β equal to the average percentage of incorrectly reproduced couples.

The resulting beta distribution represents a continuous distribution of probable true accuracy values, given the observed performance of each stochastic model. The accuracy value used to represent each model’s performance was then taken as the value at the upper bound of the 95% CI within the model’s estimated beta distribution—that is, the value at the 97.5th percentile within the distribution. For each stochastic model, this value represents an estimate of the best likely true accuracy of that model within the population in that this value represents the largest accuracy value that would not be conventionally considered significantly different from the model’s accuracy distribution.

Figure 6 presents the results of using this best likely summary to estimate the performance of stochastic models in couple simulation. With this summary method, the couple simulation method does produce correct inferences in the aspiration threshold population, given sufficient sample size. Starting at k = 200, the ATM has significantly higher simulation accuracies on average than all alternative choice algorithms. Furthermore, this does not occur merely because the best likely estimation procedure is overly generous: The ATM does not outcompete the GSA or RAM in the populations where these algorithms are the ground truth.

OPEN IN VIEWER

Figure 6.

Results of couple simulation using best likely accuracy estimation for stochastic models. (A) Random population, (B) KHM population, (C) ATM population, (D) GSA population, and (E) RAM population. Error bars represent 95% confidence intervals.

However, couple simulation does not perform well in the random and Kalick-Hamilton populations. All models performed poorly in reproducing observed couples in both of these populations, with all models averaging just M = 2.04% simulation accuracy across sample sizes in the random population and just M = 5.57% in the Kalick-Hamilton population. This is in contrast to an average simulation accuracy of M = 17.04% across models and sample sizes in the aspiration threshold population. Overall, reproducing couples that were originally produced according to either the random or Kalick-Hamilton mate choice algorithms appears to be difficult for all mate choice algorithms.

This poor performance is especially clear when k = 300 or greater, where the couple simulation method tends to have the best discrimination between models. At these sample sizes, no model achieves greater than 10% simulation accuracy in either the random or Kalick-Hamilton populations; the couple simulation method does not perform this poorly in the other three populations even when k = 500. The couple simulation method furthermore fails to yield correct inferences on average in either of these populations, with the random model and KHM producing similar performances in the random population and the ATM performing best in the Kalick-Hamilton population (although achieving just M = 5.68% accuracy, 95% CI = [5.61%, 5.74%], when k = 300). This suggests that if the true model of mate choice is as stochastic as these models are, all models will perform poorly in reproducing observed couples in couple simulation and differential accuracy between models will cease to be informative.

Similar results are obtained when analyzing the proof-of-concept models in terms of the frequency with which couple simulation produces correct inferences. Supplemental Figure 2 presents confusion matrices showing the proportion of samples in which each model produced the highest simulation accuracy across all populations and across all sample sizes. These matrices echo the average performance plots in showing that couple simulation will not frequently yield accurate inferences in the random or Kalick-Hamilton populations, regardless of the sample size. However, for the models that are sufficiently deterministic so as to produce higher overall simulation accuracies—the ATM, GSM, and RAM—discrimination is substantially better. Again, once the sample size is above k = 300, a decision heuristic that assumes the most accurate model is the true model will yield the correct inference in at least 80% of samples in each of these three populations.

One concern raised by these results is that couple simulation and the best likely estimation procedure are “overfit” to the ATM. That is, this method only yielded correct inferences for one of the three stochastic models and so may not yield correct inferences for other stochastic models either. To determine the risk of this, I created slightly stochastic versions of the GSA and RAM and tested them against one another and their deterministic counterparts in a separate set of proof-of-concept models (see Supplemental Materials Section S3). The best likely estimation procedure yielded correct inferences from couple simulation in this set of populations, with each model—stochastic or deterministic—producing the highest simulation accuracy, both on average and in the majority of samples, only in those populations where it was in fact the ground truth model of mate choice. This suggests that couple simulation’s poor performance in the random and Kalick-Hamilton populations reflects a limitation imposed by highly stochastic mate choice, but not by stochastic mate choice in general.

Lessons from the Proof-of-Concept Models

Overall, these proof-of-concept models reveal several important things about couple simulation. First, if mate choice is reasonably deterministic, as in the aspiration threshold, Gale-Shapley, and resource allocation populations, the couple simulation method can successfully reveal a population’s true model of mate choice. Applying the couple simulation method to random samples of mated couples will result in the true model of mate choice outperforming alternative models in reproducing sampled couples both on average across samples and in the majority of samples drawn. Attending to the relative simulation accuracy of alternative models can therefore provide a means of determining which models of mate choice are relatively more plausible.

Second, couple simulation cannot accurately discriminate between models of mate choice when the true model of choice is highly stochastic, as in the random and Kalick-Hamilton populations. This is a limitation of the method. However, the practical consequences of this limitation are likely to be minimal for at least two reasons. The first is theoretical: Given the incredible importance of mate choice to fitness, it is unlikely that human mate choice truly is as stochastic as the random or KHM models imply. Throughout human evolutionary history, random mating would have left the critical fitness bottlenecks of reproduction, inheritance, and parenting up to chance. Individuals who selected their mates randomly would have been severely outcompeted by individuals who were more discerning about their choice of romantic partner. For this reason, modern humans are likely endowed with psychologies that attempt to inject as much order into mate choice as possible.

This is not to say that there is no randomness in mate choice or mate choice psychology. At a minimum, for instance, the pool of potential mates available is a random function of time and place—a level of randomness that couple simulation emulates by exposing each agent to a random sample of alternatives. Mate choice psychology may also contain random elements itself; for instance, random mate pursuit as in the ATM could function to motivate exploration of a wide set of potential mates. Nonetheless, over human evolutionary history, decision psychologies that were as random as the random model and KHM would be unlikely to lead to reliably fitness-beneficial mate choice and as such would likely be replaced by mate choice psychologies capable of less random mate choices, even if not fully deterministic choices. For this reason, across populations, true models of mate choice are unlikely to be more stochastic than the couple simulation method can tolerate.

The second reason is methodological: Because all models of mate choice perform poorly when mate choice is highly stochastic, researchers can detect stochastic mate choice using couple simulation by paying careful attention to simulation accuracy across models. If a researcher were to apply couple simulation to a sample and observe very low simulation accuracies from all models, this researcher can and should infer that mate choice within their study population is substantially random—at least with respect to the features they measured. In this context, they should not attend to which of the poorly performing models performs best as these differences in accuracy are not likely to be meaningful. A hard cutoff for overall simulation accuracy is difficult to recommend, as several factors appear to affect overall accuracy, especially sample size. Keeping sample size in mind, researchers should instead be increasingly cautious in interpreting differential simulation accuracy across models as overall accuracy decreases; lower simulation accuracy across models should suggest greater randomness in the true model of mate choice, and should warrant greater caution in interpreting relative differences between alternative models.

Finally, these proof-of-concept models provide some recommendation about optimal sample sizes for applying the couple simulation method. In all populations, simulation accuracy declines with increasing sample size for all models of mate choice. Very large sample sizes would be both practically prohibitive to collect in human samples and could drive simulation accuracy low enough to prevent distinguishing between models. However, smaller sample sizes yield smaller gaps in simulation accuracy between the true model of mate choice and alternatives. A sample size close to k = 300 couples appears to provide a good trade-off between overall simulation accuracy and discrimination between models of mate choice.

Study 1

Participants in the first study were the n = 382 people (k = 191 heterosexual romantic dyads) used to generate the agent populations for the proof-of-concept models. Participants were recruited using Qualtrics’s survey panel service; the recruitment criterion was that participants were in a romantic relationship and were cohabitating at the time of participation. The target sample size was k = 200 dyads. A total of 230 dyads were initially recruited. Of these, 39 dyads were excluded for at least one participant reporting being nonheterosexual, yielding the final sample size of k = 191 heterosexual dyads. Participants ranged in age from 21 to 82 years and were M = 49.86 years old on average (SD = 14.48) and were in their relationships for Mdn = 12.08 years. The majority of participants (347, 90.83%) were married; 4 were engaged, 14 were dating seriously, and 2 were dating casually. A total of 13 participants reported their relationship status as “other”; in text descriptions, these participants mostly described their relationships as being a “domestic partner” or “living with partner but not married.” No participants indicated that they were not in a committed romantic relationship with their partners. Participants completed the survey one at a time, with the first member of the couple completing the survey in its entirety and in privacy before their partner beginning.

All participants in this sample reported their ideal mate preferences in a long-term, committed, romantic partner using a mate preference questionnaire with 20 7-point bipolar adjective scales. These asked participants to report, for example, how intelligent they desired ideal long-term mate to be on a 7-point scale ranging from “Very unintelligent” to “Very intelligent.” The questionnaire asked about preferences known from the prior literature to be universally important (e.g., kindness, intelligence, health) or universally sex differentiated (e.g., physical attractiveness, financial prospects, masculinity/femininity, age); some preferences were included as they were expected to evoke more variable responding (e.g., religiosity, sociability). In addition to reporting their ideal preferences, participants also rated themselves and their partner on the same 20 dimensions. Self and partner ratings were averaged together for each participant to form composite ratings of each participant’s own trait standing on each dimension. Self-other agreement was reasonable for these ratings, averaging r = .60 across dimensions and ranging from a low of r = .44 to a high of r = .72. Supplemental Tables 1 and 2 summarize the means and standard deviations of these preference and trait ratings and their correlations, respectively.

Trait and preference ratings were used to parameterize a population of agents based on the human participants for use in couple simulation. I generated 382 agents, each of whom was assigned the 20 preference and 20 composite trait values of one real participant. The identity of this agent’s real-world partner was also saved. This population of avatar agents then ran independently through each of the five mate choice algorithms from the proof-of-concept models; these algorithms were operationalized identically as in the proof-of-concept models. The GSA produced a matrix of paired agents; the simulation accuracy of this model was calculated as the proportion of these pairs that matched the real-world couples on which the avatar agents were based. The RAM ran for 100 model steps. Two agents were saved as a couple in this model if, by the end of the 100 model steps, both agents mutually allocated the majority of their resources to the other. Because not all agents achieved a mutual match, this resulted in 169 couples of a total 191 possible (88.48%); the simulation accuracy of the RAM was computed as the proportion of the real-world couples accurately reproduced in these model-predicted couples. Unpaired agents were included in this proportion and were considered inaccurate predictions by default.

The stochastic models—the random model, KHM, and ATM—were each run 100 times in total. Each run produced a matrix of predicted couples, and the accuracy of a model run was calculated as the proportion of real-world couples accurately reproduced in that run’s predicted couple matrix. I then calculated the average percent accuracy for each model across the 100 runs and used this to estimate each model’s best likely accuracy using the same procedure as in the proof-of-concept models.

CIs on simulation accuracy were estimated through bootstrapping. For all models, this was accomplished by sampling, with replacement, from the models’ simulation accuracies to generate 10,000 random bootstrap samples. For the deterministic models, the bootstrap procedure sampled with replacement from the model’s couplewise simulation accuracy vector—a binary vector of length n, with each element being either 1 (if that participant’s relationship was accurately reproduced) or 0 (if that participant’s relationship was inaccurately reproduced), producing 10,000 vectors of length 382. The accuracy of each of these bootstrap samples was calculated as the sum of the vector. For stochastic models, the bootstrap resampled from the couplewise accuracy vectors across the 100 runs of each model. This produced 10,000 382 by 100 resampled accuracy matrices for each stochastic model; simulation accuracy was then calculated for each bootstrap sample exactly as in the original sample. For all models, the limits of the 95% CIs were estimated as the values in the 2.5th and 97.5th percentile in the resulting vectors of 10,000 bootstrap accuracies. A separate set of proof-of-concept analyses indicate that this is a viable method of estimating CIs on simulation accuracy (see Supplemental Materials Section S4).

Couple simulation accuracy

Figure 7 presents the results of applying the couple simulation method to this sample, including the simulation accuracy for each model as well as 95% CIs. Two findings are key here. First, although the random model and KHM perform poorly, as they do in the proof-of-concept models, couple simulation in this sample is substantially above floor across models. The average simulation accuracy across all models is 24.88%, indicating that the couple simulation method can successfully reproduce real human couples and that the true model of mate choice, within at least this population, is not likely to be so stochastic as to render differential accuracy uninformative.

OPEN IN VIEWER

Figure 7.

Couple simulation results for Study 1’s sample of romantic dyads. Error bars represent bootstrapped 95% confidence intervals.

Second, and more importantly, the couple simulation method also showed discriminative simulation accuracy. Not all models performed equally well, and performance ranged from the random model reproducing just 2.38% of real-world couples up to the best performing model, the RAM, reproducing a surprising 48.69% of the real-world couples. By comparison, a prediction model that, rather than simulating mate choice, merely guessed that each participant was paired with the potential mate who best fulfills their mate preferences would accurately identify no more than 26.96% of couples. It is also worth noting that the rank order of model accuracies and the relative difference between models are qualitatively similar to the same results in the proof-of-concept population in which the RAM was the ground truth mate choice algorithm.

A paired permutation test can estimate a traditional p value comparing any two models. When applied to this sample, such a test indicates that the RAM is in fact significantly more accurate than the next best model, the GSA, p < .001. Nonetheless, this result should be interpreted cautiously, as the sample size (k = 196) is smaller than the sample size where couple simulation more accurately discriminates between alternative models (k = 300).

Relationship quality

Furthermore, a surprising finding emerged from this sample relating simulation accuracy to relationship quality. The RAM performed best at reproducing the couples in this sample; nonetheless, this model still inaccurately reproduced more than 51% of the real-world couples. Much of this inaccuracy surely stems from the fact that the RAM is not a perfect model of human mate choice. However, some of these inaccurate predictions may also reflect that some of the couples are in some way unstable. Although no relationship between simulation accuracy and relationship quality was predicted in advance, several prior hypotheses predict that variables and processes related to mate choice (e.g., preferences, CLs, or mate value discrepancies; e.g., Conroy-Beam et al., 2016; Fletcher & Simpson, 2000; Shackelford & Buss, 1997) will be related to subsequent relationship quality. These hypotheses generally propose reciprocal relationships between mate choice and relationship quality, where unfavorable mate choices motivate negative appraisals of relationships, and these negative appraisals then motivate search for more favorable partners. Couple simulation was designed to evaluate models of choice, and not to model these relationship quality appraisals. However, to the degree that couple simulation accurately simulates participant mate choices, and to the degree that choice processes do relate to subsequent relationship quality, couple simulation may nonetheless attain power to predict features of relationship quality.

To explore this possibility, I took advantage of a separate set of measures available in this sample. In addition to the preference and trait ratings described above, participants also completed a battery of relationship quality, relationship behavior, and personality measures. These measures were largely exploratory in nature and included a number of pilot measures as well as measures intended for separate projects. Here I focused analysis on those measures that most clearly tapped into overall relationship quality. This included two measures of relationship satisfaction, the Quality of Marriage Index (QMI, Norton, 1983) and the satisfaction subcomponent of the Perceived Relationship Quality Components Questionnaire (PRQC; Fletcher et al., 2000); a measure of investment and two measures of commitment, from the Investment Model Scale (Rusbult et al., 1998) and the PRQC; a measure of jealousy, the Multidimensional Jealousy Scale (Pfeiffer & Wong, 1989); a measure of attachment, the Experiences in Close Relationships Short Form (Wei et al., 2007); and a measure of romantic love, the Triangular Love Scale (Sternberg, 1997). Participants completed all these measures in random order.

To analyze these, I conducted a series of multilevel models predicting relationship quality scores from simulation accuracy. Each of these models had participants nested within dyads, with random intercept terms, to account for the nonindependence of dyadic data. For each, I also standardized the relationship quality measures prior to analysis. Simulation accuracy was a binary predictor: Each couple was either accurately reproduced by the RAM or inaccurately reproduced. Standardizing the outcome variables in these models means that the resulting b values can be interpreted much like a Cohen’s d: The b reflects the standard deviation difference between accurately and inaccurately reproduced couples.

Consistent with the possibility that the couple simulation method was capturing aspects of relationship quality, couples that were accurately reproduced by the RAM had significantly more positive scores on each measure of relationship quality. Figure 8 plots the differences across dimensions. Participants in accurately reproduced couples reported significantly higher relationship satisfaction according to the QMI (b = .46, SE = .12, p < .001, 95% CI = [.22, .71]) and the PRQC (b = .49, SE = .13, p < .001, 95% CI = [.24, .74]). These participants also reported greater investment in their relationship according to the Investment Model Scale, b = .33, SE = .12, p < .001, 95% CI = [.09, .57], as well as more commitment according to the Investment Model Scale (b = .43, SE = .13, p < .001, 95% CI = [.18, .67]) and the PRQC (b = .29, SE = .13, p = .021, 95% CI = [.05, .54]). Members of accurately reproduced couples additionally reported stronger feelings of romantic love, b = .52, SE = .13, p < .001, 95% CI = [.27, .76]. Finally, these participants reported lower levels of jealousy, b = −.44, SE = .12, p < .001, 95% CI = [−.68, −.20]; attachment avoidance, b = −.64, SE = .12, p < .001, 95% CI = [−.89, −.40]; and attachment anxiety, b = −.36, SE = .12, p = .003, 95% CI = [−.59, −.12].

OPEN IN VIEWER

Figure 8.

Differences in relationship quality in Study 1 as a function of couple simulation accuracy for the resource allocation model. For ease of comparison, the y-axis represents the percentage of maximum possible score for each relationship quality variable. Boxes reflect standard boxplots. Box edges represent the first and third quartiles (Q1 and Q3, respectively); the center line represents the median. The length of each whisker represents no more than 1.5 times the interquartile range (IQR); points are individual observations less than Q1 or greater than Q3 by more than 1.5 times the IQR.

In addition, the predictive power of simulation accuracy does not appear to be explained by any of a variety of potential predictors of relationship quality. To compare potential sources of simulation accuracy’s predictive power, I ran a series of control analyses that fit multilevel models predicting QMI satisfaction from both RAM accuracy and a control variable: one of either mate preference fulfillment, self mate value, partner mate value, partner-self mate value discrepancy, partner-potential mate value discrepancy, or relationship length. I limited these analyses to just QMI satisfaction to limit the number of models fit. Simulation accuracy remained a significant predictor of relationship satisfaction in each of these control analyses (largest p value = .026). This suggests that couple simulation’s power to predict relationship quality does not derive from couple simulation merely picking up on low-level predictors of relationship quality but rather relates to higher level features of mating markets themselves.

To see one such potential feature, we can compare true relationships and predicted relationships for those participants who were mispaired by the RAM. These participants were no more similar in overall mate value to their predicted partner than to their true partner, b = .17, SE = .13, p = .204, 95% CI = [−.09, .43]. However, predicted partners were better matches to participants’ preferences than were their real partners, b = .43, SE = .02, p < .001, 95% CI = [.26, .62]. This observation also provides some insight into why simulation accuracy declines with sample size: In larger samples, the likelihood is greater that, for each participant, the simulated mating market will contain an alternative that better fulfills their mate preferences. To the extent that couple simulation mispairs participants with more preference fulfilling alternatives, overall simulation accuracy will tend to decline in larger samples.

Finally, one concern with these results is that they may simply reflect biases shared between participant preferences and partner perceptions; for instance, perhaps participants who misperceive their partners as embodying their preferences are more likely to be reproduced accurately and are more likely to report higher relationship quality. However, all couple simulation results, for both simulation accuracy and relationship quality, persist even when participants’ own perceptions of their partner are removed from all analyses—albeit with smaller effect sizes (see Supplemental Materials Section S6).

Furthermore, not only did the RAM significantly predict each of these relationship quality dimensions, but according to the Akaike information criterion (AIC), models predicting relationship quality from resource allocation simulation accuracy were better fits to the data than models predicting the same dimensions with the GSA or the ATM. To predict relationship quality using the ATM, I estimated the best likely simulation accuracy for each couple and recoded this as a one if the best likely accuracy was above .5 and as a zero otherwise. The RAM was the superior model for predicting relationship quality for nearly all dimensions; the only exceptions to this were for PRQC commitment (ΔAIC = 2.237 relative to the ATM), relationship investment (ΔAIC = 4.50 relative to the GSA), and attachment anxiety (ΔAIC = 1.31 relative to the GSA).

Overall, these results suggest that couple simulation not only has the power to discriminate between alternative models of mate choice but also can identify happy, committed romantic relationships and models that perform better at reproducing couples overall tend to also have greater power to predict relationship quality. This set of findings is particularly surprising in that none of the models of mate choice utilize information about relationship quality: All are based exclusively on mate preferences and traits. Nonetheless, these models attain power to predict individual relationship satisfaction, commitment, investment, love, jealousy, and attachment.

Study 2

Study 1 produced several surprising findings. I therefore conducted a second dyad study with two aims: (a) to replicate the findings of Study 1, including the high on-average simulation accuracy, discrimination between models, and power to predict relationship quality and (b) given the apparent predictive power of simulation accuracy, to estimate the stability of this classification over time. Study 2 was conducted nearly identically to Study 1: Participants were cohabiting couples recruited using Qualtrics’s survey panel service. The key change was the addition of a longitudinal follow-up, completed at least 28 days after first participation. Participants in Wave 1 were n = 274 people making up k = 137 committed romantic dyads. Participants ranged in age from 24 to 89 years and were M = 48.88 years old on average (SD = 12.71). Again, nearly all participants were married at the time of participation (n = 246, 89.79%) and were in their relationships for Mdn = 17.42 years. Participants completed the same set of measures as in Study 1, with the addition of the complete PRQC, adding the intimacy, trust, passion, and love subscales. Descriptive statistics and correlations for preference and trait ratings are summarized in Supplemental Tables 1 and 2.

For Wave 2, participants were invited by email to complete the same set of measures 28 days after their initial participation. A total of n = 80 participants, making up k = 40 dyads, completed Wave 2, yielding a 70.80% attrition rate. Participants who completed both sessions of the study did not differ significantly from participants who only completed the first session in Wave 1 age, income, mate preference fulfillment, overall mate value, or relationship satisfaction according to the QMI or PRQC (ps range from .072 to .542). Participants who completed both sessions did report significantly higher feelings of romantic love at Wave 1 than people who did not complete both sessions according to both the Triangular Love Scale, t(183.95) = 2.13, p = .035, d = .26, 95% CI = [−.01, .52], and the PRQC, t(176.56) = 2.25, p = .026, d = .28, 95% CI = [.01, .54]. Participants were not limited in how long after their initial participation they could complete Wave 2; the average delay between Waves 1 and 2 was M = 31.40 days, with a range from 28 to 56 days.

Couple simulation accuracy

Participant trait and preference data were again used to parameterize agents for couple simulation. Figure 9 presents the simulation accuracy results for all participants in Waves 1 and 2. Just as in Study 1, simulation accuracy was high across the board with an average model accuracy of 25.40% in Wave 1 and 32.56% in Wave 2. Furthermore, the relative ordering of model accuracy replicated in both waves, with the RAM accurately reproducing the largest proportion of couples. Again, by comparison, a model that merely predicted each participant was paired with the potential mate who best fulfilled their preferences would accurately predict no more than 21.17% of couples in Wave 1 and 28.75% of couples in Wave 2. CIs were wider in Study 2 owing to the smaller sample size, especially for Wave 2; nonetheless, the RAM was still marginally significantly more accurate than the GSA, p = .063 in Wave 1. In Wave 2, the RAM did not significantly outperform the GSA, p = .223, and only marginally outperformed the ATM, p = .059. However, both sample sizes, with k = 137 at Wave 1 and k = 40, are well below the k = 300 sample size where discrimination between models tends to be higher. Overall, the simulation accuracy findings from Study 1 were replicated in Study 2.

OPEN IN VIEWER

Figure 9.

Couple simulation for Study 2’s sample of romantic dyads. (A) Wave 1 and (B) Wave 2. Error bars represent 95% confidence intervals.

Relationship quality

Despite the smaller sample size, Study 2 also largely replicated the relationship quality findings from Study 1. Due to power concerns, I only analyzed data from Wave 1. Figure 10 shows that, again, participants whose relationships were accurately reproduced by the RAM reported higher relationship quality across nearly all dimensions. Analysis of relationship quality proceeded identically as for Study 1, and as such, all reported b values can be interpreted equivalently to Cohen’s d values. These participants reported significantly higher relationship satisfaction according to the QMI (b = .64, SE = .15, p < .001, 95% CI = [.35, .94]) and PRQC (b = .68, SE = .15, p < .001, 95% CI = [.39, .97]). Just as in Study 1, simulation accuracy’s power in predicting QMI satisfaction was not eliminated by controlling for mate preference fulfillment, self mate value, partner mate value, partner-self mate value discrepancy, partner-potential mate value discrepancy, or relationship length (all ps < .022). Accurately reproduced couples further reported more higher commitment according to the PRQC (b = .47, SE = .15, p = .002, 95% CI = [.18, .76]) and Investment Model Scale (b = .39, SE = .14, p = .007, 95% CI = [.11, .67]); higher investment (b = .70, SE = .15, p < .001, 95% CI = [.41, .98]); higher love according to the Triangular Love Scale (b = .66, SE = .15, p < .001, 95% CI = [.37, .96]) and the PRQC (b = .56, SE = .15, p < .001, 95% CI = [.27, .85]); and higher intimacy (b = .64, SE = .15, p < .001, 95% CI = [.34, .94]), trust (b = .48, SE = .14, p < .001, 95% CI = [.20, .76]), and passion (b = .67, SE = .15, p < .001, 95% CI = [.37, .96]) according to the PRQC.

OPEN IN VIEWER

Figure 10.

Differences in relationship quality in Study 2 as a function of couple simulation accuracy for the resource allocation model. For ease of comparison, the y-axis represents percentage of maximum possible score for each relationship quality variable. Boxes reflect standard boxplots. Box edges represent the first and third quartiles (Q1 and Q3, respectively); the center line represents the median. The length of each whisker represents no more than 1.5 times the interquartile range (IQR); points are individual observations less than Q1 or greater than Q3 by more than 1.5 times the IQR.

These participants also reported significantly lower avoidant attachment than inaccurately predicted participants (b = −.53, SE = .14, p < .001, 95% CI = [−.81, −.24]). However, the difference in total jealousy was just marginally significant in Study 2 (b = −.28, SE = .15, p = .06, 95% CI = [−.56, .01]) and the difference in attachment anxiety was not significant (b = −.17, SE = .14, p = .20, 95% CI = [−.44, .09]).

As in Study 1, nearly all these relationship quality variables are generally better predicted by couple simulation accuracy according to the RAM, compared with the GSA or ATM. The only exception to this was for attachment anxiety, ΔAIC = 5.94 relative to the GSA. Attachment anxiety was significantly predicted by the GSA (b = −.38, SE = .14, p = .006, 95% CI = [−.65, −.11]) but not by the RAM. Unfortunately, unlike Study 1, the smaller sample size of Study 2 precluded analyses comparing predicted partners with true partners.

Theory Development

There are several open directions for future research that are likely to be tributary to the broader goal of stimulating theory development in human mate choice research. Some of these future directions pertain to the development of computational models that incorporate more elements of real mate choice and more realistic simulations of these elements. As argued above, computational models are formal expressions of theories; “logical machines” that convert theoretical assumptions about the nature of mate choice into predictions (Gunawardena, 2014; Smaldino, 2017). Developing computational models that make more realistic and more comprehensive assumptions thus provides a pathway to theories of mate choice that are more thorough and precise. Future research can advance theory development in the domain of human mate choice by applying couple simulation to further advance model development. Below, I detail seven specific future directions in which couple simulation could be applied to guide further theory development.

The clearest first step in this direction is to continue applying couple simulation to evaluate broader sets of computational models of human mate choice. For tractability, in developing and initially applying the couple simulation method, I have focused here on just five example models of mate choice. However, these are far from an exhaustive set. Although these models represent some of the most frequently cited models, there are alternatives within the literature (e.g., French & Kus, 2008; Knittel et al., 2011; Smaldino & Schank, 2012). Even the models considered here have alternative implementations. I considered a single, abstracted version of the ATM, most similar to that described in Todd and Miller (1999), but a variety of related versions of this model have been proposed before (e.g., Hills & Todd, 2008; Todd et al., 2005; Todd & Miller, 1999). The KHM has a variable parameter—the maximum number of dates agents go on before settling—which here I set to the value used in the original paper (Kalick & Hamilton, 1986); however, this parameter could instead be treated as a free parameter, fit to the data by finding a value that optimizes couple simulation accuracy. Similarly, within the ATM, I assumed a fixed and universal rate of threshold learning, but the degree to which each agent adjusts their threshold in response to experience could also be fit to the data as a free parameter. By testing broader sets of mate choice models in a couple simulation framework, future research will gain greater insight into what theoretical assumptions and model features yield greater simulation accuracy, expanding the parts list available for constructing novel models of choice.

Second, future models can also incorporate more realistic elements of constraint. Although all theorists agree in principle that, in mate choice, people are constrained by finite time, energy, and other resources, the exact sizes of people’s resource budgets are hard to know because these resources can be challenging to quantify. For this reason, while most of the models tested here incorporate time and resource constraint considerations, these constraints are relatively simplified and are fairly arbitrary over substantial ranges of manipulation. For instance, in Studies 1 and 2, the RAM’s simulation accuracy remains virtually unchanged even when the duration of mate choice is limited to just 10 iterations (accuracy = 46.60% and 45.46%, respectively).

More comprehensive theories of mate choice will require greater understanding of the precise temporal, energetic, financial, or computational constraints people are under when selecting romantic partners. Future research could move in this direction by attempting to measure real-world resource constraints more precisely to better inform the assumptions of future computational models of mate choice. Alternatively, and compatibly, constraint assumptions within computational models could be treated as free parameters and manipulated so as to find values that optimize simulation accuracy. In this way, through couple simulation, models and real-world data could reciprocally inform one another to yield more realistic estimates and models of constraints on mate choice.

Third, despite the name, couple simulation need not be applied to model the decisions of coupled individuals exclusively. Romantic relationships are not equally beneficial or desirable for all people (e.g., DePaulo & Morris, 2005), and many people remain unmated voluntarily. Although some of the models tested here, such as the RAM, can result in unpaired individuals, these unpaired agents at best represent involuntarily single people. If voluntarily single participants were recruited alongside mated couples, avatar agents could be simulated for these participants and their behavior modeled in simulated mating markets alongside avatars of mated people. Such mixed mating markets could be used to test models on their ability to reproduce voluntarily single people’s decisions to opt out of the mating market. Modeling the decisions of both single and mated people would yield theoretical models that more fully describe the complete mate choice decision space.

Fourth and relatedly, in the models tested here, all agents in the population pursued essentially the same mating strategy. This is not a necessary assumption, and future applications of couple simulation could test mixed models with strategic differences between agents. There is no reason to assume that all human mate choice, from all people, and under all contexts, proceeds according to the same algorithm, and so no one mate choice algorithm is guaranteed to be the best model in all contexts. Examining variability in mate choice algorithms will therefore be an essential step for future research. For instance, couple simulation could be used to test models in which different agents deploy different mate choice algorithms based on contextual factors or individual differences such as age, sex, personality, mating strategy, attachment style, relationship history, culture, or other factors.

Fifth, I have focused entirely on using couple simulation to compare mate choice algorithms. However, couple simulation could be used to evaluate models of all the tasks of mate selection. For instance, in all proof-of-concept models and in both human samples, I assumed that mate preference integration followed a Euclidean integration algorithm. This is not a necessary assumption; there are other algorithms that have been proposed as models for how preferences are integrated into evaluations of potential mates (Brandner et al., 2020; Conroy-Beam, 2018; Eastwick et al., 2014; Miller & Todd, 1998). Whereas I have held constant the model of mate preference integration and varied algorithms of choice, one could just as easily fix the algorithms of choice and vary the model of integration to observe what integration model most accurately reproduces real-world couples. Couple simulation could be pushed even further still to evaluate which preferences are important in mate choice.

Sixth, in all applications of the couple simulation method, I have additionally assumed that all individuals in the sample are equally considered as viable potential mates and, as such, all individuals have perfectly overlapping mating markets. This is not a necessary assumption and is possibly unrealistic as well. Some people are probably disqualified as potential mates prior even to the evaluation task of mate choice. This certainly includes people of the wrong gender relative to one’s sexual orientation, but could potentially also include people who are too old or too young relative to one’s age, or perhaps are just too extreme on some preference dimensions (e.g., too unkind, too unintelligent, etc.; Li et al., 2013). Indeed, people express a variety of “dealbreakers”: characteristics that, if possessed, are thought to evoke deeply negative evaluations of potential mates, regardless of other characteristics (Jonason et al., 2015). Relatedly, some have suggested that mate choice proceeds more by rejecting potential mates with negative characteristics than by pursuing potential mates with positive characteristics (Li et al., 2013; Long & Campbell, 2015). One could apply the couple simulation method to evaluate decision rules for determining the disqualification of potential mates with negative characteristics. For instance, one hypothesis testable by couple simulation is that people use a decision algorithm similar to the ATM (e.g., sequential aspiration; Miller & Todd, 1998) to cull their pool of potential mates initially and then select among the limited pool using a different mate choice algorithm such as the RAM.

Seventh and finally, all the models tested here omitted the actual mate search and perception process, simply assuming instead approximately veridical perception from all agents. Such uniformity in mate perception seems unlikely, given that participants disagreed to some extent with even their highly committed partners in their trait ratings. Future research could improve upon these models by incorporating models of the search process—and, crucially, allowing variability in perception of potential mates between agents. For instance, the mate search process could be modeled as similar to the classic multiarmed bandit problem, with the mate value of each potential mate being precisely knowable only to the degree that one has spent time actively pursuing that mate. Such a model could account for variability between people in their perceptions of potential mates as well as capture the exploration/exploitation trade-off inherent to real mate search.

The breadth of search that agents undertake is also worth some consideration. The GSA and RAM models both assume that agents are simultaneously evaluating and comparing all potential mates in their mating pool. In contrast, the KJHM and ATM models assume agents evaluate only one randomly encountered potential mate at a time. This difference in the breadth of mate search likely contributes somewhat to the difference in performance between these models. However, to test this possibility, future research must build models that vary more fully and systematically between complete and partial mate search. Intermediate models are possible as well: Agents could use random encounters to build up a pool of potential mates to which they apply their mate choice algorithms completely.

Relatedly, agents in the models tested here had all information about potential mates available to them immediately. However, in real mate choice, information about potential mates is revealed gradually, and different pieces of information are available at different points in time. Some information is available immediately—for instance, the physical attractiveness of potential mates is relatively easily observable. Other information, such as personality and intelligence, can be inferred indirectly relatively early on with moderate accuracy (Ambady & Rosenthal, 1993; Borkenau et al., 2004). But other information, especially the most critical information about how a potential mate will behave toward oneself (Lukaszewski & Roney, 2010), can only be observed by sampling observations from specific potential mates over time (Thibault & Kelley, 1959). Miller and Todd (1998) proposed a sequential model of mate choice in which different pieces of information are used to create a successive series of aspiration thresholds, with easily acquirable information used to select among potential mates at early stages of relationship development but slower emerging information setting thresholds for later relationship stages. This sequential aspiration hypothesis has not yet been modeled explicitly; however, the sequential use of information it proposes could be incorporated into models of mate choice and tested using couple simulation. Incorporating the sequential nature of real information search, and testing models that can accommodate the uncertainty it poses, will be an essential direction for future research.

Methodological Refinements

Other future directions pertain to more methodological concerns. This includes refining the couple simulation method itself and sampling data for couple simulation from a wider array of populations, under a broader set of relationship contexts, and employing more valid and informative measures. Five such methodological refinements offer clear next steps for future research.

First, these preliminary applications of couple simulation have exclusively utilized couples that are already engaged in very committed, long-term relationships. These couples have their advantages for use in couple simulation: These relationships represent the endpoints of the decision models couple simulation attempts to evaluate. Nonetheless, they have their disadvantages as well. In particular, there is no guarantee that the data collected from these participants, and used to parameterize their avatar agents, are representative of what these individuals were like when they made their initial mate choice. People of course change over time in mating relevant ways: they become older, become more educated, advance in their careers, and so on. And although mate preferences appear to be somewhat stable over time (Bredow & Hames, 2019), mate preferences do change across the lifespan (Cohen et al., 2019), and there is evidence that mate preferences also change after the initiation of relationships (Gerlach et al., 2019).

These changes are not uninteresting for understanding mate choice. To be in a long-term relationship, one must not only initiate a relationship with a partner but also maintain that relationship over time. As the existence and frequency of divorce and remarriage attest, people can and do abandon their mates in search of others—presumably at least in part due to disconnects between preferences and partners and/or the availability of more appealing outside options (Buss et al., 2017). For this reason, a mated person’s current preferences and traits are as much a part of explaining their mate selection as are their historical preferences and traits.

Nonetheless, successfully modeling initial mate choice is of clear value as well. And uncritical use of observed preference and trait values irrespective of their changes over time could lead to biased conclusions from the couple simulation method. If one has a model of how these changes occur, they can be adjusted for within the framework of couple simulation. Applying such an adjustment to the two studies reported here does result in lower simulation accuracies but does not change the central conclusions (see Supplemental Materials Section S5). Even better, the couple simulation method could be applied to simulating couples at earlier stages in the mate choice process. This could be accomplished by collecting data from newer couples or by, for instance, using couple simulation to attempt to prospectively predict the mate choices of single people.

Second and similarly, the couples used here for Studies 1 and 2 were both composed of relatively older adults in relatively long-lasting, established relationships. The median relationship lengths for these two studies were 12 and 17 years, respectively. Applying couple simulation to evaluate models of mate choice among younger people in more nascent relationships will be important as different models may perform well at explaining relationships at different stages. For instance, a model such as the ATM could describe how people decide to enter relatively casual, low-commitment, early-stage relationships, whereas the RAM could instead be describing how these early relationships develop into the highly committed relationships observed in Studies 1 and 2.

Relationship length had mixed associations with simulation accuracy in the studies tested here. A multilevel model predicting standardized relationship length from RAM simulation accuracy finds that couples do not differ in relationship length as a function of simulation accuracy in Study 1 (b = .14, p = .350, 95% CI = [−.15, .43]). However, in Study 2, couples accurately reproduced by the RAM tended to be in slightly shorter relationships than inaccurately reproduced couples (b = −.40, p = .023, 95% CI = [−.74, −.06]). That said, clearer understanding of the differences in mate choice as a function of relationship stage will come from studies able to use couple simulation to model samples composed exclusively of relatively nascent, established, or intermediate couples rather than the mixed samples used here. Applying couple simulation to test a variety of models in a variety of relationship stages and contexts will yield a clearer picture of the full trajectory of relationship development.

Third, participants sampled here were sampled from a national U.S. population, and this introduces several limitations that should be addressed in future research. First of all, these samples are of course overrepresented in psychological research in general, including mating research, and couple simulation should be used to explore both universals and variability in human mate choice across cultures (Barrett, 2020; Henrich et al., 2010). Second, the nature of mate choice in modern populations differs dramatically from the context in which our mating psychology evolved (Goetz et al., 2019). Whereas our mating psychology was designed for relatively small mating markets where any individual could expect to relatively exhaustively search and evaluate their pool of options, mate choice now takes place in anonymous mating markets on the scale of millions of people where no individual can expect to evaluate more than a fraction of the potential mates they could in principle pursue. This likely injects noise into the mate selection process and thereby noise into comparisons between models of mate choice. Third, and related to this limitation, participants in our samples do not know one another directly, and so the validity of couple simulation as applied to these samples rests on the assumption that the samples are at least representative of each participant’s actual mating market. Both of these latter limitations could be alleviated by drawing samples from smaller, more closed mating markets, either by leveraging smaller, more isolated social networks within broader social networks (e.g., college dormitories) or by applying couple simulation to data from non-WEIRD cultures with more isolated mating markets (Henrich et al., 2010). Furthermore, how mate choice differs in large, modern mating markets as opposed to smaller, more traditional mating markets is a research question couple simulation could help answer.

Fourth, another important future direction will be to determine a more effective way to evaluate stochastic models within the couple simulation method. Modeling stochastic model performance as a beta distribution allows for accurate discrimination among moderately stochastic models of mate choice. However, highly stochastic models—such as the KHM—still produce very low levels of couple simulation accuracy overall and poor discrimination between alternative models. Although it is unlikely that true human mate choice is as stochastic as the KHM, it would be ideal to have a means to demonstrate this empirically. Future researchers should continue to explore alternative means of modeling the performance of stochastic models of choice to allow discrimination between models even in populations where mate choice is highly random.

Fifth and finally, the preliminary results reported here also hint to couple simulation’s potential value as a practical tool for helping people form successful romantic relationships. Given the myriad benefits of positive romantic relationships (Antonovics & Town, 2004; Holt-Lunstad et al., 2010; Robles et al., 2014), any tool that can help guide people toward high quality romantic relationships could have important consequences for human wellbeing. Couples accurately reproduced by couple simulation—as opposed to those not accurately reproduced—reported higher relationship quality across a wide array of dimensions: satisfaction, love, commitment, investment, jealousy, attachment, and more. If this predictive power could be harnessed in longitudinal designs for prospective predictions of relationship quality, couple simulation could provide a mechanism for translating relationships theory into real-world benefits for everyday people.