Selection of hyperparameters in deep neural networks is a challenging problem due to the wide search space and emergence of various layers with specific hyperparameters. There exists an absence of consideration for the neural architecture selection of convolutional neural networks (CNNs) for spectrum sensing. Here, we develop a method using reinforcement learning and Q-learning to systematically search and evaluate various architectures for generated datasets including different signals and channels in the spectrum sensing problem. We show by extensive simulations that CNN-based detectors proposed by our developed method outperform several detectors in the literature. For the most complex dataset, the proposed approach provides 9% enhancement in accuracy at the cost of higher computational complexity. Furthermore, a novel method using multi-armed bandit model for selection of the sensing time is proposed to achieve higher throughput and accuracy while minimizing the consumed energy. The method dynamically adjusts the sensing time under the time-varying condition of the channel without prior information. We demonstrate through a simulated scenario that the proposed method improves the achieved reward by about 20% compared to the conventional policies. Consequently, this study effectively manages the selection of important hyperparameters for CNN-based detectors offering superior performance of cognitive radio network.

翻译：深度神经网络中超参数的选择是一个具有挑战性的问题，原因在于搜索空间广泛且各类具有特定超参数的神经网络层不断涌现。目前，针对频谱感知中卷积神经网络（CNN）架构选择的研究尚属空白。本文提出了一种基于强化学习和Q学习的方法，系统性地搜索并评估频谱感知问题中生成数据集（包含不同信号与信道）的多种架构。通过大量仿真实验表明，本文方法所提出的基于CNN的检测器性能优于文献中的多种检测器。对于最复杂的数据集，本方案在计算复杂度增加的情况下实现了9%的准确率提升。此外，提出了一种基于多臂老虎机模型的感知时间选择新方法，以在最小化能耗的同时实现更高的吞吐量与准确率。该方法无需先验信息，即可在信道时变条件下动态调整感知时间。通过仿真场景验证，本方法相比传统策略将获得的奖励提升了约20%。因此，本研究有效管理了CNN检测器重要超参数的选择，使认知无线电网络获得了更优性能。