Variability in individual assessment behaviour and its implications for collective decision-making

Abstract

Self-organized systems of collective behaviour have been demonstrated in a number of group-living organisms. There is, however, less research relating to how variation in individual assessments may facilitate group decision-making. Here, we investigate this using the decentralized system of collective nest choice behaviour employed by the ant Temnothorax albipennis, combining experimental results with computational modelling. In experiments, isolated workers of this species were allowed to investigate new nest sites of differing quality, and it was found that for any given nest quality, there was wide variation among individuals in the durations that they spent within each nest site. Additionally, individual workers were consistent in spending more time in nest sites of higher quality, and less time in those of lower quality. Hence, the time spent in a new nest site must have included an assessment of nest quality. As nest site visit durations (henceforth termed assessment durations) are linked to recruitment, it is possible that the variability we observed may influence the collective decision-making process of colonies. Thus, we explored this further using a computational model of nest site selection, and found that heterogeneous nest assessments conferred a number of potential benefits. Furthermore, our experiments showed that nest quality assessments were flexible, being influenced by experience of prior options. Our findings help to elucidate the potential mechanisms underlying group behaviour, and highlight the importance of heterogeneity among individuals, rather than precise calibration, in shaping collective decision-making.

1. Introduction

Collective decision-making has been studied in a wide variety of group-living organisms, especially in relation to self-organization [1,2]. However, its underlying mechanisms at the individual level are less well understood, and behavioural thresholds are frequently postulated to play a role [3,4]. Threshold-based systems form an integral part of the life-history strategies of numerous organisms, spanning a broad range of genera [5–7]. The use of such thresholds facilitates flexibility and decentralized control, and reduces the need for more cognitively taxing decision-making processes [8,9]. Colonies of social insects constitute systems that readily make use of threshold-determined behaviours, employing them to shape complex group organization [10,11]. Indeed, many problem-solving heuristics use the same rules applied by social insects at their core, and ant colonies remain one of the key sources of data for informing algorithmic architecture [12,13]. One reason for this is that within insect societies, thresholds relating to activity and decision-making allow the effective allocation of workers to various tasks, without the need for direct central control [10,14]. Accordingly, insect colonies can solve complex organizational problems, while retaining the robustness inherent to self-organized networks [15,16].

In the ant species Temnothorax albipennis, colonies undertake emigrations to new nest sites, and often have to choose among multiple options of differing quality [17]. Emigrating to the best possible nest site is of paramount importance, as nest structure may influence a colony's future success [18,19]. Previous work has shown that ants can choose between nests without directly comparing them, and thus save both time and cognitive effort [9]. The exact mechanism underpinning such decentralized nest choice remains unclear, but it has been hypothesized that workers may have considerable variation in their individual quality thresholds, each accepting a nest option only if it meets their individual ‘quality standards’ [20,21]. For this system to be effective, however, differing worker quality assessments would need to be flexibly distributed within colonies, in order to ensure the best possible emigration decisions, and avoid the pitfalls of becoming locked into suboptimal choices. Specifically, this would mean that some individuals almost always reject a nest, while others almost always accept it, thus stopping the colony as a whole from making poor choices when better options are present, and simultaneously allowing emigration choices to occur if only low-quality nests are available.

Earlier work has shown that individuals do indeed differ in their preference for accepting the same quality nest [22]. Yet, while adaptive threshold distributions have been indicated through modelling to be plausible [20,23], to date, there remains a paucity of empirical evidence to support the theory.

There is, however, a method by which the supposed quality of nests to individual workers may be measured. Chiefly, this emerges from evidence showing that when assessing a nest site, ants spend longer within the nest if they deem it to be of higher quality, while attempting to leave if it is of lower quality [9,22,24]. Furthermore, although the pre-recruitment latency of ants is shorter for higher-quality nests [20], they spend a greater proportion of this latency inside the new nest, if it is of better quality [25]. It has been hypothesized that such a trend exists, as by spending longer within a nest, workers may better contribute to a quorum threshold prior to recruiting [24,26,27]. Specifically, experiments have shown—simply by introducing workers to a new nest site by hand—that ants within a new nest will contribute to a quorum [27]. Moreover, while recruitment is the principal method of nest choice, in certain minority cases, when the current nest is destroyed and new nests are nearby, colonies are also able to choose nests without recruiting to them, instead relying on independent worker discoveries to build up a quorum within [24,28]. As such, the persistence of workers within nests, and subsequent recruitment decisions, constitute major factors in colony decision-making [21,25]. Thus, by use of a simple measure—individual visit durations—there is the potential to determine how different individuals assess the quality of various nest sites.

Here, we investigate experimentally whether workers of T. albipennis do indeed differ in their quality assessments of potential new nest sites, and, using a computational model, we evaluate the implications of this for self-organized collective decision-making.

2. Material and methods

(a). Ant colonies

Ten T. albipennis colonies were collected from the Isle of Portland, Dorset, UK, on 29 January 2016. They all contained a queen and from 59 to 280 workers. Each was assigned a random ID number, and was maintained under standard laboratory conditions, within artificial nests housed in Petri dishes [29]. All experimental work was undertaken in February 2016.

(b). Experiments

Our experimental set-up consisted of a sequence of three 100 × 100 mm Petri dishes with Fluon-coated sides, each containing a single unoccupied nest site (figure 1). These nest sites were of three different qualities, determined by their attributes; a ‘poor’-quality nest, a ‘good’-quality nest and an ‘excellent’-quality nest (figure 1). In T. albipennis, nest quality preferences have been well studied, and thus we used these to inform nest design [30]. During experiments, 16 individual scouting workers from each of 10 colonies were tested, taking 2 days per colony, for a total duration of 20 days.

Figure 1.

Nest characteristics and experimental set-up. (a) Nest characteristics: poor—clear cover, 1 mm wall height, 6 mm entrance width; good—clear cover, 2 mm wall height, 2 mm entrance width; excellent—red filter cover, 2 mm wall height, 1 mm entrance width. (b) Experimental set-up; crosses mark the introduction point of individual ants in each Petri dish. Ants were placed into the holding dish upon completion of all three assessment trials. (Online version in colour.)

At the initiation of each experiment, a colony was selected at random, and placed into a 230 × 230 mm arena, to stimulate scouting activity. We then removed a single scouting worker at random and, using a fine-tipped paint brush, placed it into one of the three Petri dishes (figure 1). Depending on its treatment group, the ant was placed into either the dish containing the ‘poor’ nest first or the dish containing the ‘excellent’ nest first. The intermediate ‘good’ nest was always second. In each colony, half of the ants were subjected to the former treatment, and half to the latter. This controlled for any effect of whether ants would encounter new nests in order of increasing or decreasing quality [31].

From this point, each ant was allowed to investigate its dish, and once it entered the empty nest site, a timer was started. For every individual, we recorded the time spent within the nest until the point that it exited, as first nest visits comprise a principal stage in nest quality assessment [32]. The ant was then moved into the next Petri dish, and allowed to assess the nest within, until it had visited all three nests. As these ants use pheromone trails to measure nest cavity area [32], the empty nests and dishes were cleaned with alcohol and detergent between visits, removing any residues. In this way, the potential for pheromonal build-up was eliminated. Once they had assessed all three nests in turn, ants were transferred into a separate holding dish, to avoid re-sampling.

This process was repeated for all 160 ants, producing three visit duration values for each individual, corresponding to the three nest qualities. Experiments provided 469 visit duration observations out of a possible total of 480, as in some cases workers failed to investigate a nest site.

(c). Computational model

To investigate how heterogeneous nest assessments may affect collective choices, we developed a computational model. The model compared the performance of colonies with heterogeneous nest acceptance, to those with homogeneous acceptance, building on established models of nest site choice with recruitment [20,23]. Assessment time is linked to nest acceptance, and hence to recruitment [22,25]. Thus, heterogeneous visit durations are likely to correspond to heterogeneous acceptance thresholds, and homogeneous visit durations to homogeneous thresholds, providing comparability with our empirical data. Our model did not explore collective nest choice in the absence of recruitment, as this occurs only when the current nest is destroyed and a new nest is nearby [24,28], and in such cases the relationship between individual and collective behaviour is linear. In the model, colonies were presented with two potential new nest sites, A and B (electronic supplementary material, figure S1). Each ant was assigned an acceptance threshold (with the addition of a random error), which was used to decide whether to recruit to a nest, or leave and search for alternatives. Ants had two potential states—assessor and recruiter—and switched from assessors to recruiters once they had accepted a nest site, subsequently recruiting their nest-mates to the chosen site (electronic supplementary material, figure S1). Once a quorum was reached in a given nest site, defined as at least half of the colony (assessors and/or recruiters) being present in a site at any single time point, the colony was assumed to have made a choice, and the simulation was ended. All simulations were run with two potential assessment distributions among workers: a homogeneous case in which all workers had the same acceptance threshold value (with the addition of a random error), and a heterogeneous case in which threshold values were normally distributed. The model was run for different scenarios in which the qualities of the potential new nest sites (A and B) were varied, and the fractions of correct choices, and times until quorum, were compared for the two assessment distributions (figure 2c–f; electronic supplementary material, figure S2). For full model parameters, see electronic supplementary material.

Figure 2.

(a,b) Frequency of worker visit durations by nest quality, in the (a) ‘poor first’ n = 238 (n_poor = 80, n_good = 80, n_excellent = 78) and (b) ‘excellent first’ n = 232 (n_poor = 75, n_good = 77, n_excellent = 80) treatment orders. (c)–(f) Comparison between the performance of colonies with homogeneous acceptance thresholds (σ = 0) and heterogeneously distributed acceptance thresholds (σ = 1) using a computational model, with a mean threshold value of 5 and the addition of a Gaussian distributed error (mean = 0, σ = 0.5). The quality of nest B (q_B) was set to a constant value, while that of nest A (q_A) was varied. (c) Fraction of correct choices when q_B = 4, (d) mean time to quorum when q_B = 4, (e) fraction of correct choices when q_B = 3, (f) mean time to quorum when q_B = 3. For full model parameters, see electronic supplementary material.

(d). Statistical analyses

To determine how worker visit durations varied with nest quality, and how treatment order influenced this, we employed a generalized linear mixed model (GLMM). This used log10 visit duration(s) as a normal response variable, nest quality and assessment order as fixed factor predictors, the interaction between the two, and worker ID nested within colony ID as random factors. Model fit was validated via evaluation of the standardized residuals.

To determine if within-individual visit durations for the three nest qualities were correlated, and if treatment order influenced this, we used a Spearman's rank correlation. All statistical analyses were performed in SPSS (release v. 23.0.0.0).

3. Results

(a). Experiments

Nest quality had a significant effect on visit duration (GLMM, effect of nest quality: F_2,464 = 75.334, p < 0.001, mean_excellent = 714.496 s, mean_good = 249.459 s, mean_poor = 126.474 s), while assessment order did not (GLMM, effect of assessment order: F_1,464 = 2.144, p = 0.144). However, the interaction between nest quality and assessment order was significant (GLMM, effect of interaction: F_2,464 = 12.373, p < 0.001; figure 2a,b), and thus we focused upon it.

Within the ‘poor first’ assessment order, differences were explained by ants spending significantly more time in the ‘excellent’ nests than in either the ‘good’ (GLMM, Bonferroni post hoc test: t = 7.724, d.f. = 464, p < 0.001, mean_excellent = 1148.154 s, mean_good = 319.154 s) or ‘poor’ nests (GLMM, Bonferroni post hoc test: t = 11.477, d.f. = 464, p < 0.001, mean_excellent = 1148.154 s, mean_poor = 93.325 s). Furthermore, ants also spent significantly more time in the ‘good’ nests than in the ‘poor’ nests (GLMM, Bonferroni post hoc test: t = 5.599, d.f. = 464, p < 0.001, mean_good = 319.154 s, mean_poor = 93.325 s).

Within the ‘excellent first’ assessment order, again, differences were accounted for by ants spending significantly more time in the ‘excellent’ nests than in either the ‘good’ (GLMM, Bonferroni post hoc test: t = 4.961, d.f. = 464, p < 0.001, mean_excellent = 444.63 s, mean_good = 194.536 s) or ‘poor’ nests (GLMM, Bonferroni post hoc test: t = 4.270, d.f. = 464, p < 0.001, mean_excellent = 444.63 s, mean_poor = 171.002 s). However, ants did not spend significantly more time in the ‘good’ nests compared with the ‘poor’ nests (GLMM, Bonferroni post hoc test: t = 0.570, d.f. = 464, p = 0.569, mean_good = 194.536 s, mean_poor = 171.002 s).

Across all nest qualities and assessment orders, worker ID had a significant random effect (worker random effect: Z = 4.397, p < 0.001), indicating that different workers were spending different amounts of time in the nests. However, colony ID did not have a significant effect (colony random effect: Z = 1.812, p < 0.070), suggesting that visit duration distributions may have been stable across colonies.

Consistent variation among workers was then confirmed by the finding that the time spent by any given ant within nests of one quality was significantly positively correlated with the time it spent within nests of other qualities. This held true for both the ‘poor first’ treatment (Spearman's rank correlation, ‘poor first’: excellent/good = 0.501, n = 78, p < 0.001, excellent/poor = 0.250, n = 78, p = 0.027, good/poor = 0.440, n = 80, p < 0.001), and ‘excellent first’ treatment (Spearman's rank correlation, ‘excellent first’: excellent/good = 0.429, n = 77, p < 0.001, excellent/poor = 0.307, n = 75, p = 0.007, good/poor = 0.362, n = 75, p = 0.001; electronic supplementary material, figure S3a,b).

(b). Computational model

In all cases, a quorum was reached more, or as, rapidly when nest acceptance thresholds were heterogeneously distributed, suggesting that heterogeneity confers a decision speed advantage (figure 2d,f). When the better quality nest was close to, or greater than, the acceptance threshold of ants, homogeneous assessments produced a higher percentage of optimum choices (figure 2c); however, when both nest qualities were significantly below the acceptance threshold of homogeneous ants, heterogeneous assessments produced a far higher percentage of correct choices (figure 2e). In cases where the new nest sites were of equal quality, percentages of correct choices were the same for both groups, but heterogeneous acceptance thresholds again provided a speed advantage (electronic supplementary material, figure S2).

4. Discussion

Our data show that individual ants do indeed spend differing periods of time within nests of different quality, and that for any given nest quality, there exists a wide distribution of visit durations across workers, spanning several orders of magnitude. In general, workers spend more time within nests as nest quality increases. However, even at the two extremes of quality, some individuals still invest a large amount of time within nests, while others invest very little. This suggests that some ants are always satisfied with a nest, while others almost never are. Furthermore, we find that the time spent by any given individual within a nest of one quality significantly and positively correlates with the time it spends within nests of other qualities. Such a trend indicates that ants are heterogeneous in their assessment of quality, yet remain consistent in their weighting of different nest options. Moreover, the order in which ants encounter new nests also has a significant influence upon the time that they spend within them, implying that prior experience may alter quality assessment. These results provide, to our knowledge, some of the first direct empirical evidence for the use of heterogeneous quality assessments in collective choice. Furthermore, our modelling results suggest that such heterogeneity may confer advantages in collective choice, especially in cases where only mediocre options are available.

In the ‘poor first’ treatment, mean visit durations significantly increased with nest quality, consistent with previous work [9,24]. However, the variation among ants within each quality represents a novel discovery, and suggests that individuals assess nests of the same quality differently. As quality increased, more individuals began to spend longer within a nest, and the distribution of visit durations became more right-skewed with increasing nest quality (figure 2a). A possible underlying mechanism is that workers were using an internal measure of quality assessment, and hence adjusting visit duration in relation to how well a nest met their own ‘standards’. This concept is strongly supported by the finding that individuals were consistent in spending either a longer or shorter average time assessing all nests, but still proportionally differed in their visit durations based on nest quality (electronic supplementary material, figure S3a). It should be noted that the role of visit duration in collective choices is likely to be proximate, as final recruitment decisions are more important for overall colony emigration dynamics [9]. However, our results provide evidence that individuals do differ widely in their assessments of nests, thus acting as a foundation for subsequent heterogeneity in recruitment.

Experimental evidence shows that as nest quality increases, a greater proportion of workers will accept it, but that some are still not satisfied and will leave to search for alternatives, indicative of varying acceptance thresholds [9,22]. Moreover, our empirical results demonstrate strong heterogeneity in assessment durations (figure 2a,b), and as assessment time predicts nest acceptance [22,25], they also suggest heterogeneous recruitment. Thus, the results from our model, which aimed to assess the potential benefits of heterogeneity in the context of recruitment, link the two findings. In simulations, colonies were able to make more rapid decisions in the presence of heterogeneous acceptance thresholds, although often at the cost of accuracy, a trend that has been observed previously in this species (figure 2c,d) [33]. Furthermore, when presented with nests of relatively low, but differing, qualities, heterogeneity was essential to allowing correct choices, while homogeneity led to an inability to distinguish between options (figure 2e,f). This is particularly relevant, as in nature, colonies often have to choose between mediocre nest sites, a trend evident from the need to construct new walls within cavities after emigrating [34]. As such, it seems likely that heterogeneity is an important factor enabling rapid emigrations, and providing flexibility when facing a range of different nest choice scenarios.

When ants encountered the ‘excellent’ nest first, although the mean time spent within nests still increased with nest quality, visit durations in the ‘poor’ and ‘good’ nests did not significantly differ from each other (figure 2b). Interestingly, this indicates that prior exposure to an ‘excellent’ nest depressed the supposed quality of ‘good’ nests, while raising the supposed quality of ‘poor’ nests, consistently across individuals (electronic supplementary material, figure S3b). It has been shown that, at the colony level, prior knowledge of options strongly influences nest assessment behaviour [35], and indeed the presence of this process at the individual level would further improve the robustness of variable quality assessments during nest choice. Specifically, if the quality assessments of individuals are influenced by experience, this would stop workers from being locked into repeatedly rejecting a nest that they deemed poor, even though it constituted the only viable option. Furthermore, as supported by previous experimental and modelling results, experiential influence would also help to ensure that even when multiple options of very high quality were present, the best nest would still be preferred [20,36].

A simplistic view of decision-making thresholds might involve swift and categorical outcomes of either rejection or acceptance. Here, by contrast, we suggest that the visit duration to a potential new nest site, a continuous variable positively associated with its quality, may underpin a probabilistic threshold rule, as observed in other biological systems [37]. Such a system would need to take many nest attributes into consideration, as is well established for these ants at a collective level [36]. Specifically, as an ant evaluates each quality metric in turn, it may take a variable amount of time before finding one that fails to meet its threshold, and so durations until nest rejection will vary. Hence, the overall visit time will increase as a function of nest site quality, effectively prolonging a worker's contribution to a quorum in a continuous fashion, prior to recruitment (figure 2a,b).

Earlier work by Pratt et al. [27,38] demonstrated that workers introduced to, or removed from, a new nest site by hand directly influenced other ants' assessments of whether a quorum threshold had been achieved. Thus, individual nest visit durations that are positively associated with overall nest quality, as demonstrated here, may influence quorum attainment at the collective level, both directly via worker presence and indirectly via recruitment decisions, favouring the choice of the highest quality nest site available. Indeed, previous studies suggest that, in some cases, emigration choices based on visit durations alone are possible, although these are in a minority, and occur absent positive feedback [21,24].

Notably, the variation we observed in assessment durations among ants is unlikely to be accounted for simply by differences in individual movement speeds. This is because the same individuals were tested for all three nest qualities, and thus interindividual speed variation would not explain the consistent increase in assessment durations seen with increasing nest quality. Consequently, we postulate that this variation is instead due to assessment differences, and so may subsequently influence recruitment likelihoods, as demonstrated by our model.

In summary, our experimental results demonstrate that a distribution of differing, but flexible, worker quality assessments does exist within T. albipennis. Modelling suggests that such heterogeneity may allow colonies to make optimal and self-organized emigration decisions, while negating the need for direct comparisons, and circumventing the dangers of suboptimal choices. Thus, while assessments are highly variable among individuals, it is from this very heterogeneity that adaptive behaviours can emerge at the level of the society.

Ethics

All applicable institutional and/or national guidelines for the care and use of animals were followed.

Data accessibility

All data are included in the electronic supplementary material, including computational model details.

Authors' contributions

T.A.O.-W. and N.R.F. conceived the study and wrote the manuscript; T.A.O.-W. conducted the experiments; N.M. developed the computational model; T.A.O.-W. and A.B.S.-F. carried out the statistical analysis. All authors gave final approval for publication.

Competing interests

We have no competing interests.

Funding

This work received no specific funding.