Implementation of a Reconfiguration Algorithm for CognitiveRadioTroy Weingart, Gary V. Yee, Douglas C. Sicker, and Dirk GrunwaldDepartment of Computer ScienceUniversity of Colorado430 UCBBoulder, Colorado 80309–0430Email: Troy.Weingart, Gary.Yee, Douglas.Sicker, [email protected]— In wireless communication systems, advances incomputer architecture and processor technology have made itpossible for functionality previously implemented in hardwareto become tunable via software. These software-defined radios(SDRs) will allow new radio devices to sense, reason, and adapt tochanges in the RF environment and/or application requirementsmaking them cognitive radios (CRs). Fully exploiting the flexibil-ity of cognitive radios, however, requires an understanding of howdifferent permutations of radio parameters impact application-specific performance metrics. For example, a CR that is notmeeting its bit loss goals could change its operating frequency toreduce the impact of interference. However, the added overheadfrom changing frequencies could result in an application failingits latency requirements.This paper describes one such method for configuring acognitive radio and demonstrates the efficacy of the techniqueon both a simulation based analysis and an in situ evaluation ona software radio platform. Our reconfiguration system quantifiesthe influence of radio parameters such as frequency agility,bit rate, and transmit power for adapting communication atthe application, medium access control, and physical layers.The method calls for exhaustively evaluating a set of CRconfigurations against a variety of performance metrics andapplying statistical processes to determine which settings willhave the most significant impact on performance. Once thisis done, the experimental results are then used to inform thedesign of an algorithm that is able to reconfigure to meetperformance goals. We show that while the method for derivingthe model functions consisten tly across both the simulation andimplementation platforms, the different evaluation platformsemphasize different parameters when controlling the radios.“The views expressed in this article are those of the authorand do not reflect the official policy or position of the UnitedStates Air Force, Department of Defense, or the U.S. Government.”I. INTRODUCTIONSoftware-defined radios have provided the means to allowmutability in both design and configuration that was onceonly possible with changes to hardware. This has allowedpost-deployment sensing, reasoning, and adapting that canrespond to changing system parameters, such as applicationrequirements (e.g., latency or throughput), spectrum policy(e.g., commands to vacate a particular frequency band), andenvirionmental conditions (e.g., noise floor).Finding an optimal solution for a given deployment, how-ever, requires an understanding of the influence of parametriccross-layer interactions among the layers within a protocolstack. Moreover, it becomes difficult to generalize these so-lutions due to the significant influence that envirionmentalchanges have on a network and the impracticality of modelingsuch factors.In this paper we attempt to address the first of these twochallenges. Our method involves using statistical methodsbased on observed metrics to develop a guided hill climbingalgorithm. We start with a measurement phase in which weexhaustively benchmark each combination of a radio’s tunableparameters and record the performance of the link in termsof latency, throughput, and packet loss. We then perform amultivariate analysis to rank the (combined) parameters thatmost influence performance. The most influential parametersare then used as inputs to a control algorithm. The controlalgorithm proceeds to vary the radio configuration basedon stated performance goals and the measured performancemetrics that are periodically communicated between nodesin a network. The adaptive algorithm that we created forour network enforced multiple performance goals, such asspecified bandwidth or latency targets, while using the leastairtime and power possible.By exhaustively evaluating each of the configurations ofa CR against a set of performance metrics, one can applya statistical process to determine what radio settings had themost significant impact on performance. It is important that theresulting performance model be able to account for the inter-actions between parameters and different performance goals.For example, while enabling error correction may improve theexperienced bit-loss, it may negatively impact latency beyonda user’s tolerance.We use techniques derived from Design of Experiments(DOE) to determine the potential interaction of input variableson output responses and develop a reconfiguration algorithmfrom the produced model [1]. While DOE has historicallybeen applied with great success in improving manufacturingprocesses, only recently has it been applied in the wirelessdomain [2], [3]. The technique is applied by first identifyingthe parameters and output metrics of a given process andthen systematically experimenting on each permutation of theinputs. The technique then relies on statistical Analysis OfVariance (ANOVA) to quantify the influence that a configura-tion (i.e., set of input variables) has on the results.In our prior work, we showed via simulation how DOEcould be used to identify how different configurations ofthe CR could improve (or degrade) network performance insimulation [2], [4], [5]. Central to our approach was the use ofthe DOE analysis to inform the design of the reconfigurationalgorithm.This paper extends our prior work by:• describing how the results of the DOE analysis canbe used to develop an effective control algorithm forcognitive radios,• demonstrating that the technique works for two differentenvironments (a simulated radio and a “real world”software radio platform),• and, we show that the technique can be extended tosupport dynamic spectrum agility.We show that the simulation and software radio platformcan minimize transmit power and utilize frequency agilityto meet application goals in the presence of an active noisesource. Our implementation, which was built on a commonoff the shelf (COTS) SDR platform, also introduces a varietyof design challenges. These issues include decisions aboutwhen and how to change configurations, how these changes arepropagated throughout the CR network, and how much timecan be spent computing a