Ant colonies outperform individuals when a sensory discrimination task is difficult but not when it is easy

Abstract

“Collective intelligence” and “wisdom of crowds” refer to situations in which groups achieve more accurate perception and better decisions than solitary agents. Whether groups outperform individuals should depend on the kind of task and its difficulty, but the nature of this relationship remains unknown. Here we show that colonies of Temnothorax ants outperform individuals for a difficult perception task but that individuals do better than groups when the task is easy. Subjects were required to choose the better of two nest sites as the quality difference was varied. For small differences, colonies were more likely than isolated ants to choose the better site, but this relationship was reversed for large differences. We explain these results using a mathematical model, which shows that positive feedback between group members effectively integrates information and sharpens the discrimination of fine differences. When the task is easier the same positive feedback can lock the colony into a suboptimal choice. These results suggest the conditions under which crowds do or do not become wise.

In many taxa, from bacteria to humans, individuals cooperate to evaluate their environment and make collective decisions (1, 2, 3, 4). By combining multiple assessments, these groups attain more precise estimates and make more accurate decisions than solitary animals (5, 6, 7). For example, when many diverse individuals independently estimate a quantity (e.g., humans guessing the number of marbles in a jar), the group’s average estimate is often close to the actual value (8–10). Similarly, groups of fish can better distinguish between stimuli than individuals (1), and pairs of homing pigeons take more efficient routes than solitary birds (11). Condorcet’s Jury Theorem and the Central Limit Theorem show that the accuracy of a group of independent decision makers improves with group size (12–16).

Despite these expectations of group advantage, many examples of collective failure have been documented (6, 17, 18). People in groups sometimes arrive at absurd decisions they would not likely make on their own (19). Collectives fail even in estimation tasks, for which they should be especially good at filtering out random errors made by individual members (10). The problem is that groups often violate the assumption of independent decision making that is central to the wisdom of crowds (1, 19). Theories assume that individual assessments are made in isolation and then integrated by a centralized process such as vote counting or averaging. Real groups instead rely on decentralized mechanisms in which interactions and positive feedback bring the group to consensus (20, 21, 22). These interactions may either improve intelligence by integrating multiple assessments or hurt it by amplifying mistakes (23). Understanding which outcome will happen requires a comparison between individual and group performance across a range of challenges.

House hunting by Temnothorax ants provides an excellent model system for this approach (24–30). Colonies are able to select the better of two nest sites, even when few if any individual ants have the opportunity to assess more than one option (24). A collective decision instead emerges from a competition between recruitment efforts at different sites (31). To facilitate this decentralized decision making, these ants use positive feedback in the form of tandem running recruitment and a quorum rule, whereby transport to a nest site accelerates when a threshold number of visiting ants is exceeded (32, 33). At the same time, if ants are experimentally isolated, they can still distinguish two sites of different quality, allowing the decision making of individuals to be compared with that of colonies (34, 35). This approach has shown that colonies can effectively compare a larger option set than individuals (35) and that they are less vulnerable to irrational preference shifts induced by decoys (34). No study, however, has determined how group advantage varies as the difficulty of the task changes.

To address this issue, we adapted methods typically used to study sensory discrimination in individual humans or animals (36–40). Temnothorax rugatulus subjects (either an individual, a small colony, or a large colony) were given a series of choices between a constant nest, with a very dim interior light level, and a comparison nest (Fig. 1). The comparison nest was always brighter than the constant nest, but its exact brightness was varied across tests to provide an array of challenges of varying difficulty. All subjects were expected to prefer the constant nest in every test, because these ants have a strong and unambiguous bias toward darker sites (26, 27). Thus, any choice of the comparison nest could be taken as a failure to discriminate the options. Discrimination ability was measured in terms of the functional relationship between the brightness of the comparison nest and the probability of choosing the constant nest.

Fig. 1.

Experimental arena for sensory discrimination tests. In each trial, subjects started in the home nest and were induced to choose between the constant and comparison sites. The constant nest was the same in all tests, with a very dim and highly favored interior light level. The comparison nest was always brighter than the constant nest, but its exact brightness varied across tests to provide an array of challenges of varying difficulty. Nests consisted of a cavity cut into a wood partition, with a glass ceiling and floor. Cavity light level was modified by adding transparent neutral-density filters to the ceiling. Numbers in parentheses indicate interior light levels.

Results

As expected, both individuals and colonies made more accurate decisions as the choice became easier (Fig. 2A). However, the shapes of the discrimination curves were quite different. We fitted separate sigmoidal response functions to the colony and individual data:

Fig. 2.

The proportion of correct choices made as a function of site illumination difference. (A) Colonies (red triangles) outperform individuals (blue circles) when the illumination difference is small (<40 lx), but individuals perform better for larger illumination differences. (B) Large colonies (green circles) perform better than small colonies (yellow triangles) over the entire range of tested illumination differences, although this difference is not statistically significant. Thick lines represent fitting to the psychophysics function specified in Eq. 1, with parameters sampled from the Bayesian posterior distribution calculated from the data. Dashed lines give 95% confidence intervals.

where x is the difference in brightness between the constant and comparison nests, α is the discrimination threshold (i.e., the smallest detectable difference in illumination), β is a scale parameter, and λ is the asymptotic level of discrimination (41). When the differences were small (i.e., difficult choices), colonies more precisely discriminated options than individuals did. Specifically, the median discrimination threshold α was 7.4 lx for colonies and 32.3 lx for individuals (Monte Carlo test for α_individual > α_colony, P = 0.0047). When the choice was easy colonies no longer outperformed individuals, as shown by the higher asymptotic performance of individuals (λ_individual = 0.93) compared with colonies (λ_colony = 0.80; Monte Carlo test for λ_colony < λ_individual, P = 0.050). Large colonies marginally outperformed small ones across all differences (Fig. 2B), but this advantage was not statistically significant (SI Appendix, Fig. S2).

To explain these patterns we further investigated the mechanisms by which individual ants make a decision. We first considered the behavior of isolated ants, whose decision-making behavior is, in comparison with that at the colony level, relatively poorly understood. By observing lone ants as they chose between two sites of different quality, we established that they make multiple visits to each site, eventually accepting one of them as a new home with a probability that depends on site quality. Furthermore, their probability of choosing the better site grows with the number of times they compare sites (i.e., visit the two sites in succession; SI Appendix). On the basis of these observations we developed a simple Markov chain model of an ant choosing between a good nest A and a mediocre nest B. Every time the ant visits a site, she accepts it with a probability p that depends both on its quality and on the number of times i she has compared A and B:

where q_A > q_B are the intrinsic qualities of the nests. Under Eq. 2 the effect of each comparison is to reduce the probability of accepting the inferior nest, while leaving acceptance probability for the better nest unchanged. This model reproduces the positive effect of visit number on the likelihood of choosing correctly (SI Appendix, Fig. S5). It also assumes that this likelihood increases faster for larger quality differences between the nests (i.e., larger differences between q_A and q_B). This feature reflects the relative ease of distinguishing sites that are very different in quality.

We assumed that a similar process applies to individual ants within intact colonies, but in combination with social information. Consistent with earlier experimental work on the collective decision making of Temnothorax ants, we incorporated the use of quorums in our model (31). When the population of ants at a nest site surpasses a threshold, commitment to that site increases significantly (32, 33). With this addition, the site acceptance probabilities now depend on the number of ants at the site as follows:

where N_A and N_B are the number of ants in the two nests, and T determines the population at which a quorum is reached and the probability of commitment increases sharply (42). The parameter c determines the relative weighting of individual assessment and social information (c = 0 for solitary ants).

When the model was used to simulate the nest discrimination experiment, the output matched our observation that both individuals and colonies become more accurate as the choice becomes easier (Fig. 3). It also reproduced the observed advantage of larger over smaller colonies across the entire range of difficulty (SI Appendix, Fig. S6). More interestingly, the model showed the distinctive crossing of discrimination curves seen in the data, with colonies doing better at difficult choices and individuals doing better for easier ones (compare Fig. 2A and Fig. 3). To further test this prediction, we gathered more data for choices involving comparison stimuli on either side of the crossing point observed in the first experiment (SI Appendix, Fig. S3). Adding these data and refitting Eq. 1 gave revised discrimination thresholds α_colony = 6.6 lx and α_individual = 30.9 lx, and saturation levels λ_colony = 0.78 and λ_individual = 0.89. These data provided even stronger support for the better performance of groups on difficult tasks (P = 0.0020) and similar results for the better performance of individuals on easy tasks (P = 0.052).

Fig. 3.

Proportion of simulated individuals and colonies (population 100) selecting the better of two nests. Colonies perform better when the quality difference between the nests is small, but individuals choose more accurately when the difference is greater. In the corresponding experiments, a nest’s quality is set by its brightness, with darker nests being more attractive. For these simulations, q_A = 0.20, and q_B varies between 0.19 and 0.001. The quorum parameter c = 1.1.

Discussion

Groups are better than individuals in making difficult decisions, but the opposite effect is found when decisions are easy. The model suggests that the reason lies in the different assessment mechanisms operating at the level of individuals and colonies. For a difficult choice, solitary ants have a relatively high probability of accepting the worse nest, because they rely on quality-dependent acceptance probabilities that differ little for similar nests. Successive comparisons cause these probabilities to diverge, but the ant is likely to make her decision before this slow process has had much effect. Whole colonies, on the other hand, do much better at difficult choices, because they use social information to accentuate the quality difference between sites. Rather than rely on individual comparisons, the colony’s choice emerges from a competition between recruitment efforts. Recruitment generates positive feedback on the number of ants at each site, with the better site slightly favored by its higher acceptance rate. The quorum rule amplifies this difference, allowing the colony to settle on the better site more frequently.

For very easy choices, the relative advantage is reversed between colonies and individuals. Acceptance probabilities now diverge quite rapidly with comparison, allowing solitary ants to make the right choice with high probability. For colonies, social information adds little benefit, and sometimes imposes a cost. This cost arises when random fluctuations yield a quorum for the inferior nest, locking the colony onto the wrong choice. In this situation, colonies get little advantage from individual comparison, even for easy choices, because the positive feedback process is quick, with most ants accepting one site without ever visiting the other one (24).

In the model, decision quality depends on the social information factor, c (SI Appendix, Fig. S7). Poor decisions can occur both when no pooling of information occurs (c = 0) and when ants rely too much on the presence of nestmates (c > 1). In the latter case, colonies get locked into their nest choices very quickly, leading to suboptimal decisions for easy problems, as occurred in our experiments. This locking on to suboptimal options through positive feedback is a recurring feature of group decision making (12), observed both in fish (1) and in trail-laying ants (23). There exists, however, an intermediate range of c values whereby colonies outperform individuals for both easy and difficult problems (SI Appendix, Fig. S7). We may conclude then that the colonies in our experiments did not achieve the maximum possible accuracy. This result may be partially explained in terms of a speed–accuracy tradeoff. The colony’s decentralized mechanism sacrifices some level of accuracy to make a more rapid decision (43–46). Our model predicts such a tradeoff for easier choices, with greater individual accuracy coming at the cost of much slower decision making (SI Appendix, Fig. S8). This is consistent with the observation that individuals take much longer to arrive at a decision than is typical for colonies (35). However, there is not always a speed–accuracy tradeoff (47). When the choice is difficult, by sharing the burden of nest assessment, colonies make decisions both faster and more accurately.

The patterns of collective and individual performance that we have observed in ants may also be relevant to many other taxa. Nonlinear responses to social information are known in many animal groups, including humans (48), fish (1) and cockroaches (49). Studies in honey bees have shown that social interactions do not always improve collective foraging, depending on the distribution of food (50, 51). Our study confirms this point and further demonstrates that these interactions sometimes even lower collective performance.

Although the wisdom of crowds has been acknowledged for decades, its failure or “madness of crowds” (52) has been recognized for even longer (17). Our work suggests that “madness” and “wisdom” are inescapably entwined. Enhanced group cognition is a clear benefit of living together (12–16), but in situations in which decision making should be straightforward there is a cost to relying on the judgment of others.

Materials and Methods

Nest Designs.

To measure decision-making performance, we offered subjects binary choices between two kinds of nest: a constant nest and a comparison nest. The constant nest was the same in all choices, with an interior light level of 1 lx. The comparison nest was always brighter than the constant nest, but its specific light level was varied across trials (7, 14, 20, 28, 39, 56, or 112 lx). Each nest was made from a balsa wood slat (2.4 mm thick) sandwiched between glass microscope slides (50 × 75 mm). A circular cavity (38-mm diameter) was cut through the middle of the slat, and a 2-mm entrance opening was cut out of the side of the nest (SI Appendix, Fig. S1). The floor of the cavity consisted of a single glass slide, whereas the roof was made from two stacked slides. Interior illumination was adjusted by placing transparent neutral-density filters (Rosco Cinegel) between the two roof slides. This design prevented ants from directly contacting the filters, which can sometimes build up an electrostatic charge that the ants find repellent. Nests were illuminated by two standard fluorescent light fixtures suspended 37 cm above the bench on which all experiments were carried out. This provided even illumination of 1,400 lx at the benchtop, as measured by a Lutron LX-101A light meter. Each fixture had a single 8,000K T-8 full-spectrum daylight bulb (All-Glass Aquarium).

Subjects.

Thirty-two colonies of T. rugatulus were used for the colony-level tests. Half of them were small (20–80 workers), and half were large (150–250 workers). An additional eight medium-sized colonies (100–130 workers) provided 16 worker ants (2 ants per colony) for the individual tests. Only a minority of workers in Temnothorax colonies participate actively in nest-site scouting and transport of nestmates and brood items, whereas the rest of the colony waits at the home nest (24, 33). To ensure that we tested only these active ants, we placed brood items outside of the colony’s nest and selected workers that attempted to retrieve them. We reasoned that ants willing to leave the nest and retrieve brood were likely to carry out similar tasks during colony emigrations (34).

Colonies were collected in the Pinal Mountains near Globe, AZ (N 33° 19.00’, N 110° 52.56’, W). All had at least one queen, with worker populations ranging from 20 to 250 and brood populations ranging from 20 to 280. We estimate that 25–100 ants (one-third of each colony’s worker population) actively participated in scouting, site assessment, and recruitment (53). Each colony was housed in a nest like those described above, but without any light filters. Each nest was kept in a plastic box (11 cm × 11 cm), the walls of which were coated with Fluon to prevent the ants from escaping. Each box was provided with a water-filled plastic tube capped with cotton and an agar-based diet that was refreshed weekly (54).

Experimental Procedure.

In each test, a subject (either an individual or a colony) was made to choose between a constant nest and a comparison nest. The comparison nest was always brighter than the constant nest, but its exact brightness varied across tests to provide an array of challenges of varying difficulty. Each subject was tested once for each of the seven comparison nests. The sequence of comparisons was varied across subjects to control for order effects. Four distinct orderings were used, with equal numbers of subjects assigned to each (SI Appendix, Fig. S1).

Tests of individuals began with placement of a home nest against one wall of a rectangular plastic arena (17.8 × 12.7 cm) with Fluon-coated walls. Two brood items were deposited just outside the nest entrance, and a single ant was placed in the arena. The ant was then given 12 h to find the nest and to move the brood inside. This procedure ensured that the subject was aware of both the home nest and the brood items. Constant and comparison nests were then placed against the arena wall opposite from the home. To induce a move to one of these nests, the roof of the home nest was removed. The ant’s choice was assayed after 12 h by recording the nest to which she had carried the brood items. Each of 16 individuals was to be tested seven times (once for each comparison nest); however, two individuals died and one lost her mark (see below) during the experiment. Thus, a total of 106 trials were conducted. In all trials, subjects successfully moved the brood items to the home nest, but 12 trials had to be excluded from further analysis because the subject failed to move to a target nest. Between trials, subjects were returned to their original colony. Each subject was marked with a unique paint drop on its gaster so that it could be identified and retrieved for subsequent trials. Before each experiment, all glass slides were washed in a commercial dishwasher, and the experimental arena was cleaned with ethanol. Balsa slats were made fresh for each experiment and never reused.

Tests of colonies followed a similar procedure. The colony in its old nest was placed next to a standard home nest in the center of the arena. The roof of the old nest was then removed to induce the colony to move into the home nest. In all cases, colonies finished migrating within 12 h. Then, target nests were introduced, and the roof of the home nest was removed to force a choice between the targets. The colony’s choice was assayed by recording the nest to which a strong majority of colony members (>90%) migrated. Each of 16 colonies was tested seven times with different comparison nests, for a total of 112 trials. In nine trials colonies split, and in one trial the colony did not move to either target nest. These 10 trials were not included in the analysis. Order effects were controlled as for the individual tests (SI Appendix, Table S1).

Markov Chain Model.

A Markov chain model was used to describe the Temnothorax house hunting process. In the model, each ant begins in the Exploring state, from which she can either find the better nest, entering state A, or the worse nest, entering state B. If the ant fails to find either nest, she continues exploring. Once at a site, she may also leave and find the other site, with some probability. With each subsequent nest visit, the ant enters state CA_i or CB_i, where i represents the number of comparisons she has made between nests. This number is incremented each time the ant goes from A to B, or from B to A. At each state the ant has a probability of accepting the nest she is currently assessing, modeled as the final states a and b. SI Appendix, Table S2 provides a summary of model states; SI Appendix, Table S3 the model’s transition probabilities; and SI Appendix, Table S4 a description of the model parameters. The flowchart in SI Appendix, Fig. S4 summarizes all possible state transitions.

The probability of an ant committing to a particular nest is a function of the nest’s intrinsic quality q, the number of times i she has compared it with the other nest, and social information mediated by a quorum rule, as given by Eq. 3. Under the quorum rule, the probability of committing to a site increases in a step-like way with site population. For the purpose of calculating these populations (N_A and N_B), we added the number of fully committed ants (state a or state b) to 20% of the assessing ants (states A and CA_i or states B and CB_i). This reflects the approximate percentage of time actually spent by those ants inside the nest, where they can contribute to another ant’s assessment of quorum attainment. A colony was considered to have chosen a nest when more than half of the ants were in the corresponding final state (a or b).