Multi-agent Reinforcement Learning for Task Allocation in Cooperative Edge Cloud Computing

Edge cloud computing has become a fundamental computation infrastructure supporting the resource-limited devices of Internet of Things (IoT). An important problem in edge cloud computing is how to allocate tasks to the servers while minimizing various costs and satisfying task requirements. Studies to date usually assume a self-interested setting where each edge/cloud server is owned by one user who tries to maximize own interests. However, with the strong development of smart communities like smart factory, the servers are usually owned by an organization like an IT corporation. This triggers the necessity for edge/cloud server cooperation to maximize team interests. Thus, in this paper, we consider a new problem called cooperative edge cloud computing where edge/cloud servers cooperate with each other to perform tasks to optimize the interests of the whole system. This problem is difficult due to some features such as 1) the tasks usually have high workloads which cannot be well performed by only one server; 2) the tasks usually have a dependency relationship; 3) edge servers are usually distributed where each server only has a partial observation. Our idea is to formulate the problem as a multiagent system, where each server is regarded as an agent who can learn to execute decision-making for task allocation based on its observation (e.g., current server status and arriving task). Then, we employee multiagent reinforcement learning methods to make agents learn from the environment by themselves without previously designed rules. Our expected impact is that our algorithm can offer significantly better attributes such as low latency and low energy consumption in the cooperative edge cloud computing.