Biological transportation networks must balance competing functional priorities. The self-organizing mechanisms used to generate such networks have inspired scalable algorithms to construct and maintain low-cost and efficient human-designed transport networks. The pheromone-based trail networks of ants have been especially valuable in this regard. Here, we use turtle ants as our focal system: In contrast to the ant species usually used as models for self-organized networks, these ants live in a spatially constrained arboreal environment where both nesting options and connecting pathways are limited. Thus, they must solve a distinct set of challenges which resemble those faced by human transport engineers constrained by existing infrastructure. Here, we ask how a turtle ant colony’s choice of which nests to include in a network may be influenced by their potential to create connections to other nests. In laboratory experiments with Cephalotes varians and Cephalotes texanus, we show that nest choice is influenced by spatial constraints, but in unexpected ways. Under one spatial configuration, colonies preferentially occupied more connected nest sites; however, under another spatial configuration, this preference disappeared. Comparing the results of these experiments to an agent-based model, we demonstrate that this apparently idiosyncratic relationship between nest connectivity and nest choice can emerge without nest preferences via a combination of self-reinforcing random movement along constrained pathways and density-dependent aggregation at nests. While this mechanism does not consistently lead to the de-novo construction of low-cost, efficient transport networks, it may be an effective way to expand a network, when coupled with processes of pruning and restructuring.

Study species
Queenright colonies of Cephalotes varians were collected from state-managed Hammock forest in Key Largo, Florida, and queenright colonies of Cephalotes texanus were collected from live oaks on private property near the city of Gonzalez, Texas. Colony collections followed the procedures of Powell (2009). In brief extensive baiting was used to locate all nests in a focal tree, and a lack of aggression among ants from different nests was used to confirm membership of nests to the same polydomous colony. The first set of experiments, performed in June to July 2017, used three colonies of C. varians (V1, V2, and V3) and three colonies of C. texanus (T1, T2, and T3). The second set of experiments, performed in October to November 2017, used four colonies of C. varians (V4, V5, V6, and V7). The colony sizes (comprising workers, soldiers, and queens) at the start of each experiment for V1 to V7 and T1 to T3 were 190, 100, 140, 184, 64, 137, 56, 43, 55, and 82, respectively. 

Experimental setup
To examine how spatial constraints affect turtle ant nest choice, we constructed arenas that had two sections with different structural features and levels of constraint (Fig 1). One section had cavities that were fully connected to one another (F), while the other section had cavities that were linearly connected in sequence (L), with greater distances between the cavities. The F section had cavities that were closer to each other with more possible pathways between them, so it was less constrained than the L section. The arenas were made with boxes (dimensions: 11 x 11 x 3.75 cm high) in a grid arrangement as shown in Figure 1. Boxes were connected with arched paper bridges which went from the floor of one box to the floor of another, over their adjacent walls. Fluon was applied to the sides of each box to prevent ants from escaping. The artificial nest cavities all had the same entrance size and volume and were made as outlined in Powell and Dornhaus (2013, Animal Behaviour). In the first set of experiments (June to July 2017), there were three nest cavities in each section for a total of six cavities while in the second set of experiments (October to November 2017), there were four nest cavities in each section for a total of eight cavities. Additional cavities were added in the second set of experiments to allow us to assess potential nest preferences in larger colonies that may be finding and occupying all of the cavities. We refer to the two sets of experiments as the limited-cavity and additional-cavity experiments, respectively. 

Nest Choice Experiments
At the beginning of each experiment, we placed original cavities containing a single turtle ant colony in the box labeled O, equidistant from the closest cavities in the F and L sections (Figure 1). We then opened the original cavities to force the ants to move and occupy new cavities in the arena. Colony movement was filmed for the first 12 hours and cavity occupation was checked 4 times a day at 8:00, 12:00, 16:00, and 20:00. Cavity occupation was measured by lifting the outer cover of the cavities and counting the number of workers inside with minimal disturbance. Both workers and soldiers moved between cavities, but workers were counted for simplicity and because they were much more numerous. The presence of brood was also noted. Turtle ants rapidly allocate workers and brood to new cavities and differentiate between different cavity properties, so both the number of workers occupying a cavity and the presence of brood can be used as indications of their preference for the cavity. The limited-cavity experiments lasted 5 days and showed that colonies occupied new cavities within the first 12 hours, with the proportions of ants in each cavity remaining stable after the second day.  Thus, the additional-cavity experiments lasted only 3 days.

Statistical analyses
For all analyses, we considered the observation in the afternoon of the third day as being representative of the colony decision. Although later observations were in some cases available, standardizing on a single observation time makes all observations more comparable, and initial observations suggested that nest choice typically did not change after three days.

For both models, cavity connectivity (whether a cavity was in section F or L) was included as a fixed effect since we were interested in the effects of structural features on nest choice. For the limited-cavity experiments, we also included species (C. varians or C. texanus) as a fixed effect, to account for consistent differences in colony size between the two species used in those experiments. For the model of worker number in the limited-cavity experiments, we also included the interaction between species and nest type, to determine whether the two species differed in the effect of structural features on nest choice. Within all models, the colony was included as a random effect to account for consistent differences between colonies, except in the case of the brood model for the additional-cavity experiments, where including it caused a singular model fit. For the Poisson models of the number of worker ants per cavity, an observation identifier (a unique identifier assigned each time occupation was counted for each cavity) was used as an additional observation-level random effect to account for overdispersion caused by the possibility that ants aggregate non-independently in cavities. We report the fixed effects along with their 95% confidence intervals, z statistics and p-values, estimated with the Wald method.

We also used randomization tests to examine whether the observed effects of structural features on cavity occupation could have been due to chance. First, we chose just to consider the observation in the afternoon of the third day as being representative of the colony decision. The number of adult ants (workers, soldiers, and queens) was averaged across all cavities in each section (F and L) for each colony, and the test statistic was calculated as the difference between the two sections, averaged across all colonies in each experiment. In each randomization, the section label for each cavity (F or L) was randomly shuffled within each colony, and the test statistic was recalculated. For each experiment, 5000 randomizations were performed to calculate the proportion of randomizations with test statistics more extreme than the observed test statistic. 

All statistical analyses were performed with R version 4.0.3; generalized linear mixed models were constructed using package lme4.

There are two different sections of each arena, with the nest type corresponding to the section it is in. Note that the nest types here labeled as "R" and "D" are referred to as "F" (fully connected section) and "L" (linearly connected section), respectively, in the accompanying paper.