Capacitated Multi-Layer Network Design with Unsplittable Demands: Polyhedra and Branch-and-Cut

A robust control-theory-based exploration strategy in deep reinforcement learning for virtual network embedding

“…While presented and validated on service placement problems, we believe that this strategy can be broadly used to solve the exploration-exploitation dilemma in learning problems. CPU range= [5,10] CPU range= [15,20] CPU range= [5,10] MC 8 iter DQN, N=1, no ctl P-DQMC, g_P=500 I-DQMC, g_I=10 PI-DQMC, g_P=200, g_I=10…”

“…To tackle this combinatorial problem, there are typically three categories of approaches in the literature: exact methods [10], approximate and heuristic methods [4] and meta-heuristics [3]. Since exact problem-solving only works for fairly small problem instances, it is often necessary to use approximations or heuristics that allow converging very rapidly to a local minimum [4].…”

“…With, for instance, p = 2 ln n n , the generated graph is in general connected (more precisely, this probability value goes to 1 as n → ∞). The requested resources (CPU and bandwidth) of Virtual Network Requests are drawn randomly following a uniform distribution from the interval [5,10]. The system operates in a dynamic manner: during each timestep, VNRs arrive to the systemwith a mean time between arrival MTBA ∈ {1, 5, 10, 20, 40} in secondsand once a VNR is processed, the next one arrives.…”

“…We consider a variation of VNRs CPU requests happening at 15k timesteps that ends at 25k timesteps. More precisely, we consider a CPU requests range of [5,10] from 0 to 15k, then it increases to [15,20] between 15k and 25k, and eventually returns to [5,10]. The evaluated R through time of the different strategies are reported in Fig.…”