(C) PLOS One
This story was originally published by PLOS One and is unaltered.
 . . . . . . . . . .



Visual social information use in collective foraging [1]

['David Mezey', 'Institute For Theoretical Biology', 'Humboldt University Berlin', 'Berlin', 'Science Of Intelligence Excellence Cluster', 'Technical University Berlin', 'Dominik Deffner', 'Center For Adaptive Rationality', 'Max Planck Institute For Human Development', 'Ralf H. J. M. Kurvers']

Date: 2024-05 

Abstract Collective dynamics emerge from individual-level decisions, yet we still poorly understand the link between individual-level decision-making processes and collective outcomes in realistic physical systems. Using collective foraging to study the key trade-off between personal and social information use, we present a mechanistic, spatially-explicit agent-based model that combines individual-level evidence accumulation of personal and (visual) social cues with particle-based movement. Under idealized conditions without physical constraints, our mechanistic framework reproduces findings from established probabilistic models, but explains how individual-level decision processes generate collective outcomes in a bottom-up way. In clustered environments, groups performed best if agents reacted strongly to social information, while in uniform environments, individualistic search was most beneficial. Incorporating different real-world physical and perceptual constraints profoundly shaped collective performance, and could even buffer maladaptive herding by facilitating self-organized exploration. Our study uncovers the mechanisms linking individual cognition to collective outcomes in human and animal foraging and paves the way for decentralized robotic applications.

Author summary Finding and collecting rewards in ever-changing environments is key for adaptive collective behavior in humans, animals and machines. We present an agent-based simulation framework to study how individuals of groups use social information during foraging together and how this social information use shapes the collective outcome through the behavior of single individuals. Our model combines models of individual decision-making of foraging agents (evidence accumulation processes) with the movement models of these individuals in space. Our results connect decisions of individuals to group dynamics and collective outcomes in realistic physical environments, highlighting the key role of the laws of real-world constraints, bringing us closer to embodied collective intelligence. Our work introduces a flexible platform to study the interplay between individual cognitive and perceptual biases, agents’ physical environment and the resulting collective dynamics and thus also paves the way for fully decentralized mobile robot applications.

Citation: Mezey D, Deffner D, Kurvers RHJM, Romanczuk P (2024) Visual social information use in collective foraging. PLoS Comput Biol 20(5): e1012087. https://doi.org/10.1371/journal.pcbi.1012087 Editor: Ricardo Martinez-Garcia, Center for Advanced Systems Understanding (CASUS), GERMANY Received: January 4, 2024; Accepted: April 17, 2024; Published: May 3, 2024 Copyright: © 2024 Mezey et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability: Data supporting the findings presented in the main text of this study are available under the DOI: https://doi.org/10.14279/depositonce-19489. Relevant code is accessible via GitHub (link: https://github.com/scioip34/ABM). A series of four explanatory videos is published on the AV-Portal of the Leibniz Information Centre for Science and Technology University Library (link: https://av.tib.eu/series/1546). Data for S1 Fig and S1 Text is provided in S1 Data. Funding: This research has been supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy – EXC 2002/1 “Science of Intelligence” – project number 390523135. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Competing interests: The authors have declared that no competing interests exist.

1 Introduction The integration of personal and social information is a key challenge for agents in human [1–4], animal [5–10] or robotic collectives [11–13]. Agents relying too heavily on personal knowledge may prevent the spread of beneficial information among group members, whereas depending too much on information provided by others may impede parallel exploration and reduce the collective performance [14]. Collective foraging (i.e., searching for and extracting spatially distributed resources together) provides an ideal testbed to study how collectives could optimally navigate this trade-off. Identifying and extracting resources from their local environment is necessary for all organisms, and while some animals forage alone, many also do so in groups with conspecifics [15]. Previous modeling work has investigated how individuals might weigh personal and social information in the context of collective foraging (e.g. [16–18]). This work has shown that relying heavily on social information is favoured in relatively rich and clustered environments, where the behavior of others provides useful cues about potential resource locations [17, 18]. If resources are more evenly distributed across the landscape, social information is less useful, as the success of others will not predict available resources in the environment. Garg et al. [18] further showed that the selective use of social information (e.g., use only information from close-by agents) is crucial for efficient collective foraging; it can mitigate the disadvantages of excessive social information use, especially in conditions where social information is prevalent (e.g., in large groups) or when individual exploration is limited. Exploring the specific drivers of foraging strategies, Monk et al. [17] showed that if resource units are hard to find through individual search (high “exploration difficulty”) but provide a large “exploitation potential”, agents should be more likely to attend to the behavior of others. Although this line of research has provided important insights into collective behavior, such models describe transitions between agents’ behavioral states probabilistically, e.g., assigning agents a fixed probability to respond to social information. Such phenomenological approaches have two key limitations: (1) They ignore the underlying decision-making process at the individual level, and (2) they do not account for physical constraints of movement and perception in collective behavior in physical space. Previous models use strong idealizing assumptions such as agents having global knowledge about the environment (or each other), agents moving in periodic grid worlds without physical constraints, or unrealistic communication regimes [18–20]. In reality, agents typically have limited cognitive and perceptual abilities. Previous models thus cannot explain how collective dynamics emerge in a bottom-up, mechanistic and spatially explicit way. They fail to explain how individual-level perception and cognition shape collective performance for real-world collective systems, being it cognitively-bounded living agents, or decentralized robotic applications [21]. Hence an approach that has individual-level decision making at its core and incorporates laws of the physical world is essential to better understand physical collective systems as well as to synthesize collective foraging behavior on swarm robotic platforms. To investigate the mechanistic basis of foraging decisions, researchers have started using evidence accumulation models (EAM) [22, 23]. This family of models, which is widely used in cognitive science and psychology to describe the temporal dynamics of decision-making [24], proposes that agents continuously gather evidence for different decision alternatives at a particular rate. This process continues until the threshold for one alternative is reached, leading the agent to make a decision [25]. Focusing on the decision when to leave a patch in search for another, researchers have shown how EAM parameters (e.g., drift rate and decision threshold) can represent different ways to integrate and weigh evidence about resource quality; thereby, such mechanistic foraging models are able to connect individual decision making to ecological models of foraging behavior [22]. More recently, such models have also been extended to social foraging scenarios where groups of animals exchange information about resource quality to decide when to leave a patch [23]. Unlike the phenomenological models described above, these models explicitly describe the individual-level decision making processes. However, they ignore the spatial dimension of foraging, do not explicitly describe agent movement and do not investigate the collective consequences of individual-level decision making. Here, we present a large-scale, mechanistic and spatially-explicit agent-based simulation environment [26] combining individual-level continuous evidence accumulation of personal and social cues and particle-based movement. Agents in our model do not have global information about the environment or explicitly share information with each other. Instead, inspired by natural collectives, they only have local information about resource quality and use visual perception—conforming to the laws of optics—to gather information about their social environment [27]. Equipped with this framework, we investigate the key trade-off between personal and social information use in a collective foraging task. After reproducing established findings from more abstract probabilistic models, we mechanistically show how individual-level decision making generates these collective outcomes. We further investigate the influence of different physical and perceptual constraints and show how incorporating such real-world limitations can greatly improve collective performance by buffering the effects of maladaptive social herding. Our freely available software framework [26], in-depth documentation as well as an interactive Playground tool with graphical user interface [28] set the stage for future research on collective foraging.

3 Discussion Complex collective behaviors in multi-agent systems emerge from the interplay of individual dynamics and social interactions. Agent-based models provide an established tool for studying the corresponding self-organization on macroscopic scales [35]. Agent-based models of collective behavior often make strongly simplifying assumptions, ignoring key limitations imposed by the laws of physics. This study introduces an agent-based model that integrates individual-level evidence accumulation of personal and visual social cues and spatially explicit movement, going beyond previous models by explaining collective dynamics as a result of cognitively realistic decision-making processes in a bottom-up way. Modeling choices between individual exploration and exploitation of social cues using two competing integrators, we systematically investigated how varying weights assigned to social information influence foraging performance in diverse resource landscapes, while comparing the influence of different real-world constraints. Although rarely stated explicitly, the reductionist approach of modelling collective behavior in the absence of realistic constraints relies on the implicit assumption that such constraints only modify higher-order emergent behaviors, without affecting overall qualitative results. Here, we directly test this assumption by systematically investigating the addition of different sensory and physical constraints. With similar idealizing assumptions (e.g., 360° visual FOV and physical overlaps between agents), our model indeed reproduces findings from abstract probabilistic models [17, 18]. However, accounting for more realistic interactions between agents generated qualitatively different results. For instance, introducing physical and perceptual constraints one after the other, we report a ‘reversal’ of the observed performance of social versus non-social agents: While collision avoidance due to finite body size of agents decreased group performance, introducing visual occlusion as a second type of constraint recovered the qualitative observations of the strongly idealized model. Thus, one should exercise great caution when deducing mechanistic explanations from idealized agent-based models. Our work highlights the importance of testing the robustness of model results with respect to model modification, accounting at least partially for real-world physical constraints. Our present analyses only scratched the surface of the large-scale simulation framework we introduce. We focused on the joint effects of social excitability, group size and physical constraints within different resource environments, holding other individual decision parameters (e.g., decision thresholds, the integration rate for personal information as well as cross-inhibition and decay weights; see Table 1) constant and assuming that patches, as well as agents all have the same quality and agents always exploit patches until depletion. Systematically investigating each of these factors offers exciting avenues for future research. For instance, with different patch qualities within the same environment or stochastic foraging returns, agents would face an additional trade-off between continued exploitation of discovered patches and sampling of different ones. Introducing heterogeneous groups of agents and changing agents’ sensitivity to personal information alongside their social excitability could uncover how both personal and social evidence-accumulation processes interact in such highly uncertain ecologies. Animal foraging is an integral part of behavioral ecology [36]. Optimality models such as the marginal value theorem [37], the diet breadth model [38] or social foraging theory [15] have provided predictions about optimal behavior given specific constraints, and empirical researchers have extensively studied (social) foraging decisions using, for instance, producer-scrounger games [39] or tracking data [40]. The cognitive mechanisms underlying (optimal) foraging decisions in naturalistic environments are still largely unknown. Recent evidence accumulation models on social patch-leaving decisions provide insights into individual decision-making but greatly simplify the decision environment and omit the spatial dynamics of mobile foragers [23]. By combining models of continuous evidence accumulation and movement, our model provides fresh insights into animal collective behavior. For instance, limited perceptual abilities such as a narrow FOV and visual occlusions might actually enhance rather than harm collective detection of resources or evasion of predators in many natural systems. Physical and sensory constraints affect the performance of social groups by effectively modulating the influence of social information and preventing maladaptive herding. Thus, the effective weight of social information has two contributing factors: A cognitive one, where individuals internally assign different weights to social versus non-social cues, and an ecological one, determined by the ability of individuals to interact with the (social) environment. This ecological factor can only partially and indirectly be controlled by individuals through their behavior. This distinction demonstrates that selective information use, a key mechanism to prevent overexploitation and avoid maladaptive herding previously shown with phenomenological models [18], can be also driven by environmental and perceptual constraints rather than solely by cognitive decision-making. By “embodying” selective social information use within the perceptual system, such apparent constraints might partly remove the need for more sophisticated social learning strategies and increase the value of social information across a wider range of organisms. Our results show that to fully understand how animals selectively use information, we need models that account for the interplay between perception, cognition, and the environment, offering insights beyond those provided by simpler phenomenological models. In the future, researchers might use controlled experiments as well as fine-grained tracking data to investigate how behavior generated by different evidence accumulation parameters maps onto optimal foraging predictions in naturalistic collectives with different perceptual abilities and movement patterns. Furthermore, future studies might investigate fundamentally different constraints such as energetic costs limiting the behavior of most living and engineered agents or constraints of other sensory modalities such as olfaction or audition. Our results also have implications for studies on human social decision making. Typical laboratory studies assume simplified choice environments and miss key features of real-world social systems where individuals continuously search for and integrate information and can only use (social) information that is currently accessible to them. Our simulation results clearly demonstrate that perceptual and physical constraints shape the spatial dynamics of social influence as well as their collective consequences. Accounting for such visual-spatial limitations, researchers have recently introduced immersive-reality approaches to study how human groups forage for rewards in different 3D resource landscapes [1, 41]. In line with predictions from our model, participants in those experiments strategically adjusted their social information use to the structure of the environment. Specifically, participants were more likely to switch to a “social relocation” state and approach successful group members in concentrated environments (i.e., few but rich patches) than in distributed ones [41] and were more likely to forage close to successful others in smooth (i.e., spatially correlated) compared to random environments [1]. Virtual foraging experiments also revealed more social information use and scrounging when participants were incentivized on the individual (vs. group) level [41], raising theoretical questions about the evolution of social information use in groups where individual and collective incentives do not (fully) align. Researchers could, for instance, use optimization algorithms (e.g., evolutionary strategies or reinforcement learning) and let the individual decision weights of our model or other behavioral strategies evolve in different resource environments (for example, see [42]). Comparing these evolved strategies to the collective optima discovered in the present study could elucidate when a group-beneficial level of social information use might evolve despite individual incentives to over-exploit patches discovered by others. Physical constraints play a vital role in engineered systems. Swarm robotic applications are highly sought after for various purposes, including search and rescue after natural disasters [43], exploring hazardous environments [44], and efficiently fighting forest fires [45]. The synthesis of such applications requires a thorough understanding of individual agents’ decision-making processes within collectives as well as their ability to adapt to and navigate in dynamic, real-world environments. Here, physical constraints cannot be ignored. Constraints naturally arise from the laws of the physical world (such as the laws of optics and embodiment) influencing the possible actions agents can take and shaping optimal behavior. Hence understanding the effects of realistic constraints on group performance is key for any synthetic real-world application. Crucially, our results suggest that researchers might be able to harness apparent constraints to engineer more robust artificial collectives. For instance, limited visual perception—in analogy to limited information sharing—might result in more adaptive swarms (see also [46]). Complementing computer simulations, future research might also use robot swarms as embodied model systems to study collective foraging behavior in natural and synthetic collectives. Such applications could utilize the unavoidable natural constraints on robots or their environments to obtain a minimal set of realistic assumptions for further modelling work. In parallel, agent-based models facilitate the development of efficient swarm robot platforms and help to better understand and overcome the non-beneficial effects of these assumptions. Using local information is a type of environmental constraint that comes from the limited perceptual and cognitive resources of embodied agents. Interestingly, the use of local information provides a possible solution to potentially catastrophic global information outages in robotic applications: when global information is shared across group members (such as GPS, or wireless communication) tampering or fully removing sources of such information can result in the full collapse of the system. In contrast, indirect communication such as visual perception provides an alternative that is more robust against such attempts. In principle, relying only on vision allows also for cooperation of different types of robots with incompatible communication systems. Our model can serve as a basis to understand how indirect information sharing compares to direct centralized strategies and how both influence the behavior of engineered systems. In sum, our study elucidates the mechanisms linking individual-level cognitive processes and physical constraints to collective dynamics in physically (more) realistic collectives. This not only helps to explain how collective dynamics unfold in naturalistic animal and human groups, but also paves the way for the development of novel swarm-robotic applications that, out of necessity, must overcome—but, according to our results, might also harness—the limitations and affordances of the physical world.

[END]
---
[1] Url: https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1012087

Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/