Cognitive and Adaptive Equalizer Implementation in GNU Radio

Machine learning (ML) is now being applied to wireless communications to optimize existing communication techniques. GNU Radio, an open-source developmental toolkit used to model communication systems, however, has seen few machine learning techniques implemented in its library. Real time simulators, like GNU Radio, are critical in verifying the performance of communication systems. One aspect of communication theory that can be improved by machine learning is adaptive equalization. Specifically, machine learning can be used to adjust the type of adaptive equalizer algorithm as well as the structure of an equalizer (i.e. taps, step size, filter type, etc.) to determine optimal equalizer parameters for current channel conditions. This process is referred to as cognitive equalization. The objective of this research is to demonstrate how cognitive equalization algorithms can be implemented in GNU Radio along with channel model blocks as tools for further validate the cognitive equalizers. This is valuable for analyzing modern equalization techniques in various channel conditions and communication scenarios. When a signal is passed from transmitter to receiver, the channel effects on the signal can be drastic. This is especially apparent in high frequency (HF) communications (3 MHz to 30 MHz) where the effects are caused by reflections of the transmitted signal across the ionosphere. One consequence caused by the ionosphere is that the signal will take multiple paths when arriving at the receiver. Due to the various propagation paths, several components of the signal will arrive at the receiver at different times, causing multipath. Furthermore, signal fading can be observed when multiple paths are taken by the signal during transmission and the arriving signal is summed at the receiver and cancellation occurs. Channels will often introduce noise to the signal and cause inter-symbol interference (ISI), which will create errors in signal recovery. Inter-symbol interference, multipath, and fading, will result in a corrupted signal which will be difficult for the receiver to interpret. This corrupted signal can be reconstructed at the receiver with the aid of an equalizer. Equalization can be used to optimize the signal recovery in various communication systems by removing the impairments caused by the channel from the received signal. The equalizers implemented in this work utilize a least mean square (LMS) algorithm to update the associated tap values, making them a form of adaptive equalizers. Both linear and non-linear adaptive equalizers were implemented in GNU Radio, and the implementation is discussed further in this work. An additional enhancement that can be made to improve the effectiveness of the equalizers is to utilize reinforcement learning algorithms to vary the structure of the equalizer. The reinforcement learning algorithm will update the structure based on the incoming signal’s channel effects in attempt to lower the error rates. The use of a reinforcement learning algorithm in changing an attribute of a communication system based on the channel conditions is referred to as a cognitive engine. These cognitive engines can change the actual structure of the equalizer by changing the number of parameters, such as the number of taps and step size. Alternatively, the adaptive algorithms utilized within the equalizer only change the coefficient value of the taps. The benefit of implementing cognitive engines in GNU Radio is it enables the optimal equalizer parameters to be learned for a specific set of channel conditions in real time. Additionally, the implementation of these cognitive communication systems sets the stage for future development with the incorporation of real hardware. This will allow us to conduct real-time tests with true over-the-air experiments. This work details the implementation of three cognitive engines and three different channel models within GNU Radio. The contributions of this thesis include the development and deployment of three LMS adaptive equalizers into the GNU Radio library. This development included various validation tests conducted to guarantee the functionality of these equalizers. These tests included analyzing the convergence trends posed by the LMS algorithm for the tap coefficient values and the changing bit error rate (BER) for several SNR values. Contributions related to the design of the Watterson channel model have also been included in this work. Following the implementation of the adaptive equalizers, we then deployed a cognitive engine block into GNU Radio, which utilizes the epsilon greedy RL algorithm. This algorithm was tested prior to implementation and further validated with the previously implemented adaptive equalizers. The CE and equalizers were then tested together to ensure that the CE does in fact improve the performance of the equalizers by changing the structure of each equalizer based on the current channel conditions.