RESUMO
Reinforcement Learning (RL) comes with the promise of automating network management. However, due to its trial-and-error learning approach, model-based RL (MBRL) is not applicable in some network management scenarios. This paper explores the potential of using Automated Planning (AP) to achieve this MBRL in the functional areas of network management. In addition, a comparison of several integration strategies of AP and RL is depicted. We also describe an architecture that realizes a cognitive management control loop by combining AP and RL. Our experiments evaluate on a simulated environment evidence that the combination proposed improves model-free RL but demonstrates lower performance than Deep RL regarding the reward and convergence time metrics. Nonetheless, AP-based MBRL is useful when the prediction model needs to be understood and when the high computational complexity of Deep RL can not be used.
