Institutional Repository
Technical University of Crete
EN  |  EL



My Space

Collaboration pattern detection in hedonic cooperative games with externalities

Troullinos Dimitrios

Full record

Year 2019
Type of Item Diploma Work
Bibliographic Citation Dimitrios Troullinos, "Collaboration pattern detection in hedonic cooperative games with externalities", Diploma Work, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2019
Appears in Collections


Cooperative games model the formation of coalitions of rational agents that come together to gain some form of utility which they would have otherwise been unable to collect by acting alone. Hedonic games, then, constitute the class of cooperative games that models agents with hedonic preferences, that is, agents who have preferences over their very coalitional membership, i.e. the identities of others in their coalition. Thus, an agent’s utility in such settings mirrors the satisfaction yielded from its assembled coalition. Now, cooperative games with externalities, or in partition function form, consider that agent utility is influenced by the partition of the agents space, i.e., the set of all disjoint coalitions currently in place. Existing studies, however, have not so far addressed hedonic games with externalities.At the same time, uncertainty is prevalent in most realistic cooperative game environments, and hence intra-agent collaboration under uncertainty is a topic widely studied. However, uncertainty in hedonic game settings has received only limited attention in the literature to date.Against this background, in this diploma thesis we first extend the formal definition of two well-known classes of hedonic games, namely additively separable hedonic games and boolean hedonic games, to partition function form. Then, we combine the aforementioned paradigms, and focus on agents in hedonic games with externalities who are unaware of their own preferences over partitions. We demonstrate how to extract these hidden preferences by employing well-established supervised learning methods—namely linear regression, linear regression with basis functions, and feed forward neural networks—and adapting them to the problem at hand. In the process, we make use of an evaluation metric specifically designed to evaluate the prediction accuracy of machine learning methods used to infer the underlying hedonic preferences over partitions. In addition, we show how an agent can use Gaussian mixture models to generate sets of potentially satisfactory partitions to propose in multi-agent negotiations. Finally, we put forward two novel coalition formation protocols that engage agents with hidden and conflicting preferences; and which are designed with the aim of maximizing social welfare, without the presence of a centralized entity or the ability to share information among agents.

Available Files