Scientific defense of Mr. Jochen Fink on Friday, July 22, 2022
The title of the doctoral thesis is: " Fixed Point Algorithms and Superiorization in Communication Systems"
The doctoral examination board consists of the following members:
Chair: Prof. Guiseppe Caire (TU Berlin)
Prof. Slawomir Stanczak (TU Berlin)
Prof. Nikolaos Sidiropoulos (University of Virginia, USA)
Prof. Isao Yamada (Tokyo Institute of Technology, Japan)
Supervisor: Dr. Renato L.G. Cavalcante (Fraunhofer HHI)
Abstract— This work studies fixed point algorithms and superiorization in wireless communication systems. Modern wireless systems perform a great variety of signal processing tasks, including channel estimation, precoding, combining, signal detection, and peak-to-average power ratio (PAPR) reduction. The growing demand for mobile data traffic calls for systems with higher bandwidths, larger antenna arrays, and the capability to serve an increasing number of devices. As a result, the dimensions of various optimization problems arising in wireless networks are growing continuously, which increases the computational cost of algorithmic solutions. Hence, scalable algorithms with low complexity are vital to meet both real-time requirements and power budgets. For various types of roblems, finding optimal solutions can be too computationally demanding in practice. In this case, it is desirable to strike a balance between performance and complexity. Ranging between feasibility seeking and constrained optimization, the superiorization methodology is a promising means of achieving this trade-off. It relies on the concept of bounded perturbation resilience of an iterative algorithm. A feasibility-seeking fixed point algorithm is said to be bounded perturbation resilient if its convergence to a fixed point can be guaranteed even if certain perturbations are added to the iterate in each step. In this case, the superiorization methodology can be used to define a sequence of perturbations leading to a reduced (not necessarily minimal) value of a given objective function. Compared to exact constrained minimization, superiorization often results in a lower computational cost. In this thesis, we investigate the bounded perturbation resilience of several algorithmic frameworks, including the well-known projections onto convex sets (POCS) algorithm, the adaptive projected subgradient method (APSM), and certain extrapolated projection methods. By doing so, we enable their use as basic algorithms for superiorization. We propose an algorithm for the nonconvex multi-group multicast beamforming problem based on a perturbed POCS algorithm. The proposed perturbations simultaneously reduce two objective functions, one of which is nonconvex. Then we harness the bounded perturbation resilience of the APSM by proposing an algorithm for detection in multiple-input multipleoutput (MIMO) systems based on a superiorized APSM. We also devise a deep unfolded version of this algorithm,in which the design parameters are learned using a stochastic gradient descent method. Moreover,we propose online algorithms for estimating and tracking time-varying channels with hybrid- beamforming architectures based on an APSM. The proposed channel estimation algorithms can compensate for random delay and phase variations in wideband channels. Furthermore, we devise a data-driven analog combining policy. Finally, we propose extrapolated projection methods for PAPR reduction. We devise perturbations that aim at incorporating a nonconvex constraint set, which allows the simultaneous use of certain subcarriers for channel estimation and peak cancellation. Simulations at the end of each chapter show that the proposed methods can outperform state-of-the-art techniques, while often resulting in a reduced computational cost.