Offline Learning of Nash Stable Coalition Structures with Possibly Overlapping Coalitions

Researchers have developed a new offline learning algorithm that can infer the preferences of selfish agents from a fixed dataset of past interactions to form stable, possibly overlapping coalitions. The method efficiently recovers a Nash-stable partition where no agent can unilaterally improve their utility. It works under two types of utility feedback and achieves sample-efficient learning with optimality guarantees, proven through extensive experiments to converge to a low approximation of Nash stability.