Institutional Repository
Technical University of Crete
Technical University of Crete
EN
|
EL
| URI: | http://purl.tuc.gr/dl/dias/E1A2B4AB-2F8D-4005-A019-7106B1183C2D | ||
| Year | 2023 | ||
| Type of Item | Diploma Work | ||
| License |
|
||
| Bibliographic Citation | Gerasimos Koresis, "Deep reinforcement learning for overlapping coalition formation", Diploma Work, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2023 https://doi.org/10.26233/heallink.tuc.98296 | ||
| Appears in Collections |
This thesis delves into the dynamic landscape of Overlapping Coalition Formation(OCF), within multiagent systems, leveraging the power of Deep Reinforcement Learning (DRL), to navigate uncertainties inherent in cooperative interactions.The central inquiry revolves around addressing the uncertainty regarding thedegrees of cooperation (DoC) among agent types, which arguably determine the effectiveness of coalitions formed by the agents.The study unfolds in multiple dimensions. First, an exploration of RL and DRLtechniques is undertaken, emphasizing their application to the intricate challenges posed by OCF scenarios. The core of the investigation lies in deciphering the evolving dynamics of agent interactions, with a particular focus on the uncertain nature of cooperation values represented by the DoC.In response to this uncertainty, the study integrates Graph Neural Networks(GNN) into the DRL framework. In particular, our thesis details the synergisticintegration of DRL (specifically, Deep Q-Networks - DQN) and GNNs (specifically,Graph Attention Networks - GAT), showcasing their collective capacity to adaptto the ever-changing uncertain cooperation landscape. Our experimental evaluation results underscore the efficacy of this hybrid approach in enhancing sequential coalition formation strategies under uncertainty.We explored several variants of our DRL+GNNs approach, with our simulationresults suggesting the intertwining of DQN with GAT updates of the DoC at the change of the proposer to be the most beneficial one.Finally, our work in this thesis takes the initial steps to tackle the scalabilitychallenges inherent in this multiagent domain, and lays the groundwork for future refinements and extensions.
Copyright © DIAS 2013
