Reconfiguration in Network Slicing—Optimizing the Profit and Performance

Network Slice Reconfiguration by Exploiting Deep Reinforcement Learning With Large Action Space

“…The resource for embedding the network slice is the sum of resources consumed by all flows in the network slice. To define the embedding cost, we assume the pricing functions for computational and bandwidth resources are ' c (•) and ' b (•), respectively [14]. Accordingly, the cost of embedding the network slice, i.e., the embedding cost is defined as:…”

“…Slice reconfiguration itself incurs certain resource consumption because it may cause various overheads such as signaling and retransmission overhead. Thus, the reconfiguration resource is quantified as a function of the state difference of a slice before and after reconfiguration [14]. The state of a network slice is reflected by the node mapping variable x and link mapping variable z.…”

“…3) Since NSRP is indeed a sequential decision problem, we use the long-term average cost as the optimization objective in (14). Constraint (1) to (2) ensure that one VNF of a flow is served by at most one physical node.…”

“…The authors of [13] formulated the network slicing problem as a mixed binary linear programming and proposed the penalty successive upper bound minimization algorithm. Wang et al [14] modeled the network slice reconfiguration as a profit optimization problem and addressed it with a L 1norm approximation method to maximize the profit of slice customers while reducing the reconfiguration cost.…”

“…Note that we use subscript t to denote the time throughout the paper. Different from the works [13], [14], [33] which allow flow splitting through physical nodes and/or physical links, we assume that each flow is mapped exactly onto one physical path in order to avoid coordination overhead caused…”