Zhou Qi, Chen You-Wei, Shen Shuyi, Kong Yiming, Xu Mu, Zhang Junwen, Chang Gee-Kung
Opt Lett. 2019 Sep 1;44(17):4347-4350. doi: 10.1364/OL.44.004347.
We propose and experimentally demonstrate a proactive real-time interference avoidance scheme using SARSA reinforcement learning (RL) in a millimeter-wave over a fiber mobile fronthaul system. The RL consists of three core factors, including state, action, and reward. The state is defined as a discretized value from the center frequency, the left, right, and center sub-EVM of the signal. We use five actions to shift the signal frequency in the proposed scheme, which is -20, -10, 0, 10, and 20 MHz, for the RL agent to choose the action to avoid the dynamic interference. For the agent to learn from the experience, the reward is defined as the log value of BER difference between the past and the present state. The RL-based approach is an online learning algorithm, which can learn in real time based on environmental feedback. Besides, the agent can learn from past experience owing to the Q-value table, which makes it act intelligently when facing a similar situation again. We verify the capability of the proposed scheme under both fixed and dynamic interference scenarios. The agent demonstrates an efficient intelligent mechanism to avoid the interference, which provides a promising solution for proactive interference mitigation in the 5 G mobile fronthaul network.
我们提出并通过实验证明了一种在毫米波光纤移动前传系统中使用SARSA强化学习（RL）的主动实时干扰避免方案。强化学习由三个核心因素组成，包括状态、动作和奖励。状态被定义为来自信号中心频率、信号左、右和中心子误差向量幅度（EVM）的离散值。在所提出的方案中，我们使用五个动作来移动信号频率，即-20、-10、0、10和20兆赫兹，供强化学习智能体选择动作以避免动态干扰。为了让智能体从经验中学习，奖励被定义为过去状态和当前状态之间误码率（BER）差异的对数。基于强化学习的方法是一种在线学习算法，它可以根据环境反馈实时学习。此外，由于Q值表，智能体可以从过去的经验中学习，这使得它在再次面对类似情况时能够智能地行动。我们在固定和动态干扰场景下验证了所提出方案的能力。智能体展示了一种有效的智能机制来避免干扰，这为5G移动前传网络中的主动干扰缓解提供了一个有前景的解决方案。
