Toward comprehensive models of naturalistic cooperation and competition in primates

About This Grant

Project Summary The flexible ability to work together for mutual benefits while competing against others for limited resources is a hallmark of advanced social cognition. Cooperative and competitive interactions are highly dynamic and complex. However, studying the precise behavioral mechanisms of these interactions has been challenging. This is partly due to the fact that the standard animal models in lab studies do not reliably cooperate. Moreover, typical studies do not include multidimensional behavioral measurements that are essential to understand such complex interactions. Therefore, there is a need to investigate complex social interactions in a species whose social structure strongly depends on both cooperation and competition, while tracking multiple action-based and internal state-related variables to obtain a comprehensive understanding of social behavior. The marmoset is an excellent species for studying complex social interactions grounded in context-dependent cooperative and competitive tendencies within their natural ethology. The first major goal of this proposal is to simultaneously and continuously collect multidimensional biobehavioral measurements, both action-based and internal state-based, during naturalistic cooperative and competitive interactions between freely moving marmosets. We aim to understand the functional and directionally causal dependencies of these biobehavioral variables in cooperative and competitive behaviors. The second major goal is to build comprehensive and empirically testable generative models of primate social interaction and to validate our understanding iteratively between the models and the experiments. We will use multiple modeling approaches to exploit their strengths: multi-agent reinforcement learning with recurrent neural networks will be used to learn complex patterns for prediction, and the structure and inputs to these models will be informed by dynamic Bayesian networks to increase the interpretability of the models. We will use an embodied agent-based framework, with the recurrent neural networks driving musculoskeletal models of marmosets, to better model cooperative and competitive interactions of nonhuman primates. With the comprehensive biobehavioral data and the musculoskeletal model, we will build generalizable models of primate social interaction via a multi-level constraints-based framework. Finally, we will validate the generative models of primate social interaction by inducing multiple types of in silico environmental and task manipulations that are designed to predictably alter social strategies and carrying out those experiments in vivo that significantly alter the resulting social strategy. Overall, we aim to provide the most comprehensive understanding of primate social interactions to date, along with novel generative models of such behavior.