Designing scalable and energy-efficient routing protocols for IoT low power networks is a particularly challenging problem. The IETF ROLL Working Group has defined and standardized an IPv6 routing protocol for IoT low power networks called RPL (IPv6 Routing Protocol for Low-Power and Lossy Networks) [1]. This protocol builds and maintains dynamic routes among network devices based on various objective functions (OFs) that exploit different network metrics for parent node selection (e.g., ETX-based [2], Energy-based [3]), etc.). With such OFs, RPL organizes the network topology as a Destination Oriented Directed Acyclic Graph (DODAG). However, the performance of RPL may be affected by frequent network topology changes, which may be caused by different factors like node battery depletion, link quality degradation, etc. Indeed, in such situations, the OF functions do not guarantee optimal maintenance of the RPL tree. To address this issue, this paper describes how Supervised Learning can be leveraged to improve RPL performance and energy efficiency by mitigating RPL DODAG instability when the network conditions, used by the RPL's OF functions, change frequently. We use an offline supervised learning to provide the optimal value of the transmission range (the maximal distance to which a node can send its data to another one) that mitigates the instability of the RPL network, and hence minimizes the energy consumption. The preliminary simulation results show that our proposal can improve network performance and increase network lifetime.