Safe Multi-Agent Deep Reinforcement Learning for Dynamic Virtual Network Allocation

Cooperative Multi-Agent Deep Reinforcement Learning for Dynamic Virtual Network Allocation With Traffic Fluctuations

“…There have been several studies on dynamic VN allocation (RL-based [7]- [12] and heuristic [13], [14]).…”

“…With the breakthrough of DRL in human-level control applications [16], the DRL-based VNE algorithm has been increasingly studied. Studies have mentioned the problem of the RL-based approach: the candidate actions of VN allocation exponentially increase as the number of nodes and the number of links increase [10]- [12]. Dolati et al [10] attempted to shrink the action space of the VNE problem to provide sufficient flexibility for exploring different VN mappings while retaining the efficiency of the learning process.…”

“…Yan et al [11] decompose a VNE process into a sequence of virtual node embedding to shrink the action space, and the learning agent only focuses on one virtual node of the current VN request at every single step. We previously proposed a MADRL-based dynamic VN allocation method [12]. We divided the VN demands into groups, and each agent is prepared for each group, which can shrink the action space per agent.…”

“…Several studies have addressed this problem [7]- [14]. Specifically, reinforcement learning (RL) [15], [16] has been focused on as a solution [7]- [12]. RL solves the decision problem of what action an agent should take by observing the current state within a certain environment.…”

“…Therefore, we introduce a cooperative element in which several agents jointly optimize a single reward through centralized training and decentralized execution, which can improve the efficiency of VN allocation. Previous methods [10]- [12] avoid the problem of exponentially increasing action space by restricting the agents' actions at each step or compressing the number of actions by handcrafted features (see Sec. II for details).…”