Unveiling Collaborative Group Identities in Social Synthetic Environments from Interaction Data

One of the most fundamental characteristics of human societies is the ability to gather in groups and manifest group behaviours. This social property has implications on how individuals interact and cooperate with each other. The study of group behaviours is not only fundamental for better understanding phenomena such as social conflicts, but it can also contribute to the understanding of how the environment affects their evolution. The work presented in this dissertation aims to provide a computational framework capable of inferring the presence of group structures, and consequently assign group identities, to populations of socially driven individuals, by solely analysing the ongoing levels of cooperation of the interactions. Our group modelling framework is intended to be used in computer-mediated interaction scenarios, for simplicity called social synthetic environments, which can be effectively used to simulate aspects of real-life, yet by maintaining a customisable level of control of the phenomena under investigation. Examples of social synthetic environments are theoretical games and cooperative computer games. The proposed framework is composed of two pipelined modules. The first one, namely cooperation modelling, is responsible for monitoring the ongoing interactions, evaluating their levels of cooperation, and maintaining up-to-date estimates of the whole society's complex network. The second module, namely group identity detection, leverages on the information held by the cooperation network and partitions it into groups, so that the within-group cooperation is maximized and the between-group cooperation is minimised. In this dissertation we investigated a plethora of computational methods for all the tasks and modules of the framework. More specifically (1) we relied on the so-called one-to-many approach to gather and evaluate the ongoing interactions, in accordance with we observe how one individual treats many others under the same interaction protocol, though these interactions may have not occurred at the same time (2) we considered two methods for the evaluation of the interactions into normalised cooperation values (3) we considered three cooperation network update rules, two of which widely used in reinforcement learning (4) we relied on evolutionary computation for the network partitioning task. With respect to the latter, we examined three possible chromosome representations, two evolutionary approaches, namely single population and "sealed" niching, and four mechanisms for sealed niching activation. We conducted thorough empirical investigations based on a benchmark experiment in social dilemmas where the impact of induced group on cooperation and reciprocity has been studied. From the authentic data of that study we extracted the reported behaviours that were most suitable for our investigation, and also used them to derive artificial societies of believable agents. The results obtained showed promising inferential performance of our group modelling framework. This dissertation successfully leverages on the hypothesis that it is possible to infer the existence of group structures by solely focusing on the ongoing interactions. The limitations of the approach and possible strategies to overcome them are also proposed.