Distributed Multi-Agent Learning for Service Function Chain Partial Offloading at the Edge

Multi-Access Edge Computing (MEC) along with "learning at the edge" brings unique opportunities for enhancing the utilization of resources in the next generation wireless networks. Using Network Function Virtualization (NFV), Service Function Chains (SFCs), a set of ordered virtual network functions (VNFs), can be deployed within the MEC infrastructure. The user equipment (UEs) can offload VNFs with intense computational load to the MEC servers with rich storage and computation resources. In this paper, we address the problem of partial offloading of a chain of services where each VNF of the SFC request can be either performed locally or offloaded onto a MEC server. The objective is to concurrently minimize the long-term cost of the UEs which is given in terms of both delay and energy consumption. This problem is highly complex and calls for distributed multi-agent learning techniques. We formulate the problem as a distributed multi-agent reinforcement learning problem and use double deep Q-network (DDQN) algorithm to solve it. Our simulation results show that the proposed DDQN-based solution has comparable results to an exhaustive search algorithm.