Learning the Stable Allocations of Stochastic Cooperative Games

ICollaboration among multiple agents is essential in many environments. However, when there is uncertainty and incomplete information, collaboration can be difficult. For instance, agents may not know the potential or preferences of other agents, leading to uncertain outcomes in cooperative tasks.
Therefore, to determine the optimal coalitions and allocate rewards to promote future collaboration, a process of trial and error is necessary. That is, agents can form temporary agreements based on their current knowledge and update it as they learn more through collaboration. However, the presence of learning among agents can pose two typical challenges.
First, this trial-and-error process can be expensive in real-world scenarios. Agents need to choose agreements that allow them to gain sufficient information while keeping the cost of temporary agreements low. This can lead to the exploration-exploitation dilemma, where agents need to balance exploring new possibilities and exploiting current knowledge to make optimal decisions.
Second, learning and collaboration can complicate the dynamic of the system, making it challenging to reach a stable agreement. Therefore, a learning protocol that stabilizes the dynamic of the agreement is essential.
In this project, we aim to develop methods to address those two problems in the context of cooperative games. Overall, addressing these two problems is crucial in achieving successful cooperative multi-agent learning.
Alignment with EPSRC research themes: It is very well aligned with Artificial Intelligence and Robotics. Finally, it is within the scope of Information and Communication Technologies theme.