A Predictive Model for Cognitive Radio

Such a model is needed to optimize performance in a software-defined radio.

A computational model that predicts effects of changing operational parameters of a cognitive radio has been developed as part of continuing research on cognitive / software - defined radio (C/SDR) data-communication networks. The term “cognitive radio,” “software-defined radio,” or “smart radio” denotes a radio transmitter, receiver, or transceiver, (1) much of the functionality of which is implemented in software and (2) that is able to reason about its configuration on the basis of requirements and its environment.

C/SDR has been made feasible by advances in digital data processing and communication electronics that have resulted in faster, smaller, and cheaper electronic devices: Emerging dataprocessor technology has enabled radio systems that traditionally have been implemented in custom silicon to migrate to general-purpose processors. C/SDR radio systems can, potentially, make more efficient use of the available radio-frequency (RF) spectrum and adapt to a wide range of protocols and environments. One of the key benefits of C/SDR is the ability to change operational parameters in response to changes in requirements and in the RF environment. A predictive computational model — perhaps such as the one reported here — is needed to enable a C/SDR to dynamically modify its own configuration in order to improve the performance of the overall communication system of which it is a part.

Given a set of possible system configurations and environmental conditions, a suitable model should be able to predict which configuration enables satisfaction of requirements, which could include a specified throughput or latency. Such prediction is expected to be part of a continuous, repetitive process: In a typical envisioned application (1) the model would be used to configure a software-defined radio; (2) that radio would be used for communication, and data on the achieved throughput or latency and other aspects of performance would be recorded; then (3) the recorded data would be used as input by a prediction mechanism to derive a new predictive model.

Thus far, in the development of a suitable predictive model, it has been found that multilinear regression submodels are useful for predicting bandwidth (and, hence, throughput). The use of multilinear regression submodels for this purpose is most easily understood in terms of, and has many characteristics in common with, design of experiments (DOE), which provides for performing structured experiments to explore the parameter space. In the experiments, the input parameters are permuted and results of the experiments are analyzed with respect to statistical significance.

It also has been found that the use of multilinear regression submodels is not sufficient for predicting latency. A regression tree, which is similar to a decision tree, was found to be better suited for this purpose. In a regression tree, predictions are made by “splitting” categories of data within nodes in such a way that they maximally reduce the variance. The figure shows an example of a regression tree. In using the tree, one would single out those leaves that satisfy one’s latency requirements. The set of configurations for satisfying these requirements could be determined by tracing from the affected leaf nodes back to the root. [In this example, two leaves (5.25 and 5.38 ms) are candidates for satisfying the requirement for latency of less than 6 ms. For 5.38ms, the data rate would be 2 Mb/s and the frame size would be medium.] This process of tracing would be repeated for all affected leaf nodes to obtain the set of configurations that would satisfy the requirements. The regression-tree submodel of latency could be combined with the multilinear regression submodel of throughput (or with another such submodel for satisfying a different requirement such as minimizing transmitter power) to reduce the set of candidate configurations to those that best satisfy overall communication requirements.

This work was done by Troy Weingart, Douglas C. Sicker, and Dirk Grunwald of the University of Colorado for the Air Force Research Laboratory. For further information, download the free white paper at under the Software category. AFRL-0002

This Brief includes a Technical Support Package (TSP).

A Predictive Model for Cognitive Radio (reference AFRL-0002) is currently available for download from the TSP library.

Please Login at the top of the page to download.