- Volumen: 7
- Fecha: 01 enero 2019
- Páginas: 123341-123354
- ISSN: 21693536
- Tipo de fuente: Revista
- DOI: 10.1109/ACCESS.2019.2938084
- Tipo de documento: Artículo
- Editorial: Institute of Electrical and Electronics Engineers Inc.
© 2013 IEEE.The IoT is the cornerstone of many innovating processes such as those behind Smart Cities and Smart Industries. As more and more wireless IoT devices are deployed, a newer, more congestion-resilient communication infrastructure is required to absorb the traffic from the 50 billion IoT nodes expected by the year 2020. Although 5G is said to be a key technology for the future IoT, it is not a silver bullet. Therefore, providing nodes with different Radio Access Technologies (RAT) would allow them to reap the various benefits offered by each RAT. However, the process of determining which technology should be used at any given time should not be based on uninformed intuition, but on mathematically educated choices. By making use of the mathematical framework of Reinforcement Learning, we have allowed IoT nodes to learn from previous real world data in order to derive optimal RAT-selection policies. These policies, which are implemented as Artificial Neural Networks (ANN), maximize a predefined reward closely related to classic throughput, while maintaining power consumption and operational costs below a certain limit. To allow hardware-constrained IoT nodes to use these ANNs, we have proposed the application of a quantization technique that reduces computation and memory requirements and have validated it by its implementation in a real IoT device. Finally, to evaluate the proposal, we have simulated a network of 1000 devices deployed in the city of Chicago. The obtained results are compared to those achieved with other intuitive policies to further highlight the benefits of our proposal.