Due to the rapid growth of new wireless communication services and applications, much attention has been directed to frequency spectrum resources and the way they are regulated. Considering that the radio spectrum is a natural limited resource, supporting the ever increasing demands for higher capacity and higher data rates for diverse sets of users, services and applications is a challenging task which requires innovative technologies capable of providing new ways of efficiently exploiting the available radio spectrum. Consequently, dynamic spectrum access (DSA) has been proposed as a replacement for static spectrum allocation policies. The DSA is implemented in three modes including interweave, overlay and underlay mode [1].

The key enabling technology for DSA is cognitive radio (CR), which is among the core prominent technologies for the next generation of wireless communication systems. Unlike conventional radio which is restricted to only operate in designated spectrum bands, a CR has the capability to operate in different spectrum bands owing to its ability in sensing, understanding its wireless environment, learning from past experiences and proactively changing the transmission parameters as needed. These features for CR are provided by an intelligent software package called the cognitive engine (CE). In general, the CE manages radio resources to accomplish cognitive functionalities and allocates and adapts the radio resources to optimize the performance of the network. Cognitive functionality of the CE can be achieved by leveraging machine learning techniques. Therefore, this thesis explores the application of two machine learning techniques in enabling the cognition capability of CE. The two considered machine learning techniques are neural network-based supervised learning and reinforcement learning. Specifically, this thesis develops resource allocation algorithms that leverage the use of machine learning techniques to find the solution to the resource allocation problem for heterogeneous underlay cognitive radio networks (CRNs). The proposed algorithms are evaluated under extensive simulation runs.

The first resource allocation algorithm uses a neural network-based learning paradigm to present a fully autonomous and distributed underlay DSA scheme where each CR operates based on predicting its transmission effect on a primary network (PN). The scheme is based on a CE with an artificial neural network that predicts the adaptive modulation and coding configuration for the primary link nearest to a transmitting CR, without exchanging information between primary and secondary networks. By managing the effect of the secondary network (SN) on the primary network, the presented technique maintains the relative average throughput change in the primary network within a prescribed maximum value, while also finding transmit settings for the CRs that result in throughput as large as allowed by the primary network interference limit.

The second resource allocation algorithm uses reinforcement learning and aims at distributively maximizing the average quality of experience (QoE) across transmission of CRs with different types of traffic while satisfying a primary network interference constraint. To best satisfy the QoE requirements of the delay-sensitive type of traffics, a cross-layer resource allocation algorithm is derived and its performance is compared against a physical-layer algorithm in terms of meeting end-to-end traffic delay constraints. Moreover, to accelerate the learning performance of the presented algorithms, the idea of transfer learning is integrated. The philosophy behind transfer learning is to allow well-established and expert cognitive agents (i.e. base stations or mobile stations in the context of wireless communications) to teach newly activated and naive agents. Exchange of learned information is used to improve the learning performance of a distributed CR network. This thesis further identifies the best practices to transfer knowledge between CRs so as to reduce the communication overhead.

The investigations in this thesis propose a novel technique which is able to accurately predict the modulation scheme and channel coding rate used in a primary link without the need to exchange information between the two networks (e.g. access to feedback channels), while succeeding in the main goal of determining the transmit power of the CRs such that the interference they create remains below the maximum threshold that the primary network can sustain with minimal effect on the average throughput. The investigations in this thesis also provide a physical-layer as well as a cross-layer machine learning-based algorithms to address the challenge of resource allocation in underlay cognitive radio networks, resulting in better learning performance and reduced communication overhead.

Publication Date


Document Type


Student Type


Degree Name

Engineering (Ph.D.)

Department, Program, or Center

Engineering (KGCOE)


Andres Kwasinski

Advisor/Committee Member

Raymond Ptucha

Advisor/Committee Member

Drew Maywar


RIT – Main Campus