Performance comparison of learning techniques for intelligent channel assignment in Cognitive Wireless Sensor Networks

Abstract

With the increasing number of devices sharing the 2.4 GHz ISM band, coexistence problem becomes one of the major issues experienced by Wireless Sensor Networks (WSN). Cognitive Wireless Sensor Networks (CWSNs) has been proposed in order to achieve reliable and efficient communication via spectrum awareness and intelligent adaption. The learning and decision making technique is one of the core competences of such system. In this work, there machine-learning techniques under the umbrella of Reinforcement Learning (RL) including GPOMDP, Episodic-Reinforcement, and True Policy Gradient are implemented for our proposed learning and decision making engine of CWSN. Simulation model has been developed and used for the investigation and the results are obtained for performance comparison in terms of prediction accuracy and WSN system performance. From this study, True Policy Gradient offers better prediction accuracy in comparison with the other two techniques. As results, CWSN implementing True Policy Gradient offers lowest packet delay under interference environment. Keywords assignment, Gradient cognitive wireless sensor networks, channel GPOMDP, Episodic-Reinforcement, True Policy I.

Cognitive Wireless Sensor Networks: Intelligent Channel Assignment

“…When comparing with other estimations i.e. Episodic Reinforcement and GPOMDP, the expected reward map of the True Policy Gradient provides the best fittings than other methods since the suitable values are in the same level when comparing with other estimations [12]. Note that the maps for Episodic Reinforcement and GPOMDP can be seen in our previous paper [12], which also shows the simulation results for the comparison of prediction accuracy.…”

“…The result from a traditional system is also given (labelled as IEEE 802.15.4). From our previous work [12], which focuses on the selection of suitable learning approach, results from the study show that True Policy Gradient outperforms the other two techniques with 94% prediction accuracy. However, the proposed system includes virtual channels environment classification step, which further prevents wrong decision to be implemented.…”