Enhancing IEEE 802.15.4 Access Mechanism with Machine Learning Arslan Musaddiq, Tariq Rahim, Dong-Seong Kim ICT Convergence Research Center, Kumoh National Institute of Technology, Gumi-si, South Korea. (arslan, dskim)@kumoh.ac.kr, tariq@ieee.org Abstract—The Internet of Things (IoT) network consists of resource-constrained tiny devices. An efficient channel access mechanism for densely deployed IoT devices operating in a lossy environment is one of the major challenges for future IoT networks. The IoT nodes using IEEE 802.15.4 MAC protocol increase the backoff exponent (BE) during the channel sensing period. This blind increase of BE and contention window (CW) before frame transmission affects the network performance. Therefore, in this paper, we propose to use machine learning such as a reinforcement learning (RL) mechanism to handle channel access mechanisms efficiently. The proposed mechanism is evaluated using Contiki 3.0 Cooja simulations. The simulation results indicate that the proposed RL-based mechanism enhances the network performance. Keywords—IoT; MAC; IEEE 802.15.4; channel access I. INTRODUCTION Wireless networks have become increasingly popular particularly the Internet of Things (IoT) networks have vast application areas. IoT network composed of constrained- resource devices that transmit data in a lossy environment. The reliable transmission using advanced digital coding techniques in the physical (PHY) and medium access control (MAC) plays an important role in successful IoT communication. MAC layer particularly affects system-level aspects such as throughput, reliability, fairness, and so on [1]. MAC layer is mainly responsible to provide coordination among wireless nodes for channel contention. Random access mechanism is especially used for low-cost devices because it does not require a centralized coordinator and resources can be allocated in a distributed manner. The carrier-sensing multiple access with collision avoidance (CSMA/CA) protocol is the most widely used multiple access technique. In CSMA/CA, a station continuously senses the channel for contention, and defer for a random time interval if it detects an ongoing transmission to avoid collisions [2]. The IEEE has defined two standards based on CSMA/CA protocol. The IEEE 802.11 standardized MAC layer for the wireless local network (WLAN) uses distributed coordinated function (DCF). In the DCF mechanism, the station first senses the channel. If the channel is busy, it waits for a random backoff time to sense the channel again. If it is found idle, it waits for a short duration known as DCF interframe space (DIFS). Then, it selects a contention window (CW) between a random number 0 to 2  −1 . Where BE is backoff exponent. The value of BE starts from 0 and increments by 1 each time there is a collision. The maximum backoff stage value is 5. If the channel is empty the station transmits the request to send (RTS) packet and waits for a short duration known as short interframe space (SIFS) to receive clear to send (CTS). After receiving CTS, it transmits a frame. If the acknowledgement (ACK) is received the transmission is successful, if no ACK is received the station increment BE counter by 1. If the retransmission attempt is less than the maximum allowable attempts the STA retransmits the frame. The traditional IEEE 802.11 standard does not have any energy minimization mechanisms which is necessary for scarce resource IoT devices [3]. Another IEEE standard based on CSMA/CA is IEEE 802.15.4 [4]. The IEEE 802.15.4 standard is one of the key enabling technologies for devices that require low-power, high reliability, low cost, and low computation. The IEEE 802.15.4- based IoT devices have been deployed massively in numerous application areas such as smart cities, smart industries, or smart healthcare. The IoT network is made up of a number of IoT sensors and a sink node connected in the form of a graph. The data from various IoT sensors are propagated to the sink node in a multi-hop fashion. With each transmission, the device consumes valuable resources. For example, the IoT device's radio interface is a fundamental source for energy utilization. The maximum energy is consumed during the transmission (Tx) and reception (Rx) phases. Even in the idle state a considerable amount of energy is consumed as CPU is continuously reading the data to measure the channel status [5]. In IEEE 802.11 DCF mechanism node continuously senses the channel during BE period. In IEEE 802.15.4, the channel is sensed only once at the end of BE period. In this way, a node can activate sleep mode for power saving during BE period. The one-attempt sensing does not show major performance benefits in terms of energy-saving rather it fails to handle collision if network size increases. Whenever there is a collision the BE is incremented by 1. If the frame is successfully transmitted the BE is initialized to its minimum value. We can utilize machine learning such as the reinforcement learning (RL) mechanism to create Q-values for BE decisions. In this way, we can intelligently select BE and CW size before frame transmission. In this paper, we are utilizing Q-learning which is one of the RL mechanisms to optimize the channel access mechanism of IEEE 802.15.4. 210 978-1-7281-6476-2/21/$31.00 ©2021 IEEE ICUFN 2021