Research

First and foremost, the foragers' population will be identified with the network age approach, which proved to be a better task allocation predictor than biological age. Then, the foragers' population will be divided into multiple classes based on the personalities of the group members. Personality is here understood as a measure of how an individual performs any task, and can be the result of both the individual's genetic composition and personal learned experience. The identification of groups with different behaviours will be extracted by looking at the bees' movements before, during and after communication dance events. The appreciation of known personality traits from the literature will be key to support the identification of relevant observable behavioural traits to be extracted from data and the subsequent classification into target personality groups.

Data analysis starts by detecting dancing events in the database by means of a suitably trained deep convolutional neural network (DCNN), identifying bees with their role as dancers, followers, or onlookers, which naturally correlates with bees owning different information and goals. Then, additional features will be extracted from the bee trajectory, such as dance intensity for dancers or the number of observed waggle runs for followers. Pre and post-dance trajectory analysis will be used to infer dancers' urge to share information and followers' eagerness to forage after following a dance. Finally, records from foraging experiments will add valuable information such as knowledge of visited food sources. A dedicated filter on the data must be developed for each chosen feature to identify the bee and the corresponding behaviour.

As bees are labelled with unique identifiers, it will be possible to associate features to each of them to uncover their personality class. To this end, dimensionality reduction techniques will be used to obtain low-dimensional embeddings for each bee and later determine the personalities' groups through hierarchical clustering or other machine learning (ML) approaches.

The project exploits symbolic Transfer Entropy (TE) to quantify information flow between foragers. Following this approach, different symbolic representations of the same data will be considered to capture parallel information flows within the same behaviour. The analysis will have a spatiotemporal focus around dancing events, whereby bees share information about known foraging sites. For each dance, the trajectories of the dancer and of all foragers located in her vicinity—whether they actively follow the dance or not—will be considered. Trajectories will be subsampled before encoding to capture the time interval in which the sender information best predicts the receiver’s behaviour. Although the dancer’s behaviour is expected to be more informative about the followers’ behaviours, TE will be measured in both directions to also search for any effect of the foragers’ behaviours on the dancer. This counter-intuitive information transfer may reveal feedbacks that dancers receive from (the absence of) active foragers, which may be relevant to determine the dancing intensity and duration. TE from empirical trajectories will be compared with artificial datasets created by pairing independent time series to eliminate any spurious correlation. Finally, canonical correlation analysis will determine how well the foragers’ personality embeddings explain their TE performance.

To further explore the effects of heterogeneity onto the collective behaviour of swarms, a series of experiments replicating relevant aspects of the foraging behaviour of honeybees, including different personalities observed in the population, will be designed. This will allow to finely control the experimental conditions and the levels of heterogeneity in the forager population, providing valuable insights on how adaptive responses are brought forth. The experiment requires robots to forage from multiple artificial resources in the most efficient way. The quality of such resources will change over time, and the robots must adapt to such a changing landscape. Robots will be able to move around and recruit each other through direct communication (hence simulating dance interactions) to increase the colony's foraging efficiency. A homogeneous swarm of robots is deployed first, and the parameters are optimised to maximise foraging efficiency. Subsequently, heterogeneities in the robot movement and communication behaviour will be introduced, following the experimental observations from phases 1 and 2. For instance, if a relevant information transfer is detected from followers to dancers in modulating the dance behaviour according to their personality, the robot behaviour will be similarly biased so that communication will be more or less likely depending on the individual personality and the social contingencies. With such a system, it is possible to analyse different proportions of foragers' phenotypes and diverse resource landscapes, allowing to test a wide set of combinations.