Abstract: With the development of wireless mobile communication, the contradiction between users' proliferating content demands and the limited wireless network resources is increasing. The use of Device-to-Device (D2D) communication to realize the sharing of cached contents between edge nodes can improve user experience and reduce the burden of traffic on the core network. This paper models the collaboration caching problem as a cooperative game considering the factors of interaction costs and individual rationality to optimize the system utility with limited cache space. According to whether the utility between nodes can be transferred, we discuss the cooperative game in two cases. Under the transferable utility (TU) game, the conditions for nodes to form a stable grand coalition are derived, and it is proved that the coalition has the nature of nuclear nonempty when the coalition cost of nodes satisfy certain conditions. For non-transferable utility (NTU) game, the rational nodes cannot ensure the formation of a stable grand coalition, and the number of formable coalitions increases dramatically with the number of users. Therefore, a deep reinforcement learning-based coalition formation algorithm is proposed to ensure the formation of stable coalitions within a limited time. Theoretical analysis and simulation results show that the proposed algorithm can converge to a Nash-stable optimal solution or asymptotically optimal solution, which outperforms other comparison algorithms.

Key words: node collaboration, content sharing, cooperative game, deep reinforcement learning