Machine Learning in Wireless System Design
Abstract
Abstract:The optimal design of communication networks has traditionally relied on mathematical models that describe the transmissio"n process, signal propagation, receiver noise, network traffic, network topology, and many othercomponents of the system that affec""t the end-to-end signal transmission and reception. However, there are cases where this approach fails, in particular when the mathe""matical models for one or more of the system componentsare highly complex, hard to estimate, poorly understood, don~t well-capture"" the underlying physics of the system, or don~t lend themselves to computationally-efficient algorithms. In these scenarios, a compl"etely new approach tosystem design may be beneficial. The proposed research addresses this problem by using machine learning (ML) f"or elements of wireless system design. In this approach, instead of using mathematical models for designing aparticular component o""f a communication system (e.g., a detection algorithm or a resource allocation algorithm), tools from ML such as deep learning and r"einforcement learning are used to learn and refine the design directly from data.The training data that is used in this ML approach" can be generated through models, simulations, experimental measurements, or field measurements. In some cases, such as designing de""tection algorithms, the training processis performed off-line and, once trained, the models are used on-line as part of the communi""cation network. In other cases, such as training a resource allocation algorithm, the whole network may go through an initial traini""ng phaseafter deployment, and then once operational the network keeps track of the changes (e.g., using techniques from reinforceme"nt learning). Our goal in this work is to consider four problems where the ML approach holds thepromise to provide a better design than the current state-of-the-art approaches: signal detection over channels when the channel state information (CSI) or channel mo"dels are unknown, joint source-channel coding of text, images, andvideo data, routing in wireless networks, and resource allocation" (power and spectrum) in wireless networks. Our proposed research seeks to determine if ML techniques can improve upon existing methods for certain aspectsof wireless system design. Examples of previous work where ML tools have been applied to design problems in" communication systems include multiuser detection in code-division multiple-access (CDMA) systems, decoding oflinear codes, design"" of new modulation and demodulation schemes, detection and channel decoding, and estimating channel model parameters. The approach t"aken in these previous works has been to use ML to improve on thedesign for one component of the communication system based on knowledge of the underlying model associated with that design. There are several distinguishing features of our framework in comparison" with prior work. First, when a good model for a component of the system does not exist, we use ML and deep learning to obtain the"" model directly from the data. Second, we develop new NN architectures tailored to the specific design problem weare trying to solv"e. Our preliminary results indicate that the proposed approach can be used effectively in two of the design problems we investigate" in the proposed research: to train novel signal detection algorithms using NNs withoutknowledge of the underlying channel model, a"nd to design joint source-channel encoders and decoders for text data The proposed research has the potential to significantly impact both fundamental communications and networkscience as well as the Navy~s communication networks. The work on NN detection algorithms will enable enhanced capabilities in radios and underwater acoustic communication used by the Navy. Since the proposed detector" does notrequire CSI or channel models, there will be no need for CSI estimation, which improves data rate. Moreover, the detector"" will adapt to the changing ch
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 26, 2018
- Source ID
- N000141812191
Entities
People
- Andrea Goldsmith
Organizations
- Office of Naval Research
- Stanford University
- United States Navy