A Routing Algorithm based on Semi-supervised Learning for Cognitive Radio...

A Routing Algorithm based on Semi-supervised Learning for CognitiveRadio Sensor Networks

Zilong Jin1, Donghai Guan1, Jinsung Cho1 and Ben Lee2

1 Dept. of Computer Engineering, Kyung Hee University, 446-701, Yongin, Korea2 School of Electrical Engineering and Computer Science, Oregon State University, 97331, Corvallis, OR, U.S.A.

fjinzilong, [email protected], [email protected], [email protected]

Keywords: Machine Learning, Semi-supervised Learning, Routing Algorithm, Cognitive Radio Sensor Networks.

Abstract: In Cognitive Radio Sensor Networks (CRSNs), the cognitive radio technology enables sensor nodes to occupylicensed bands in a opportunistic manner and provides advantages in terms of spectrum utilization and systemthroughput. This paper proposes a routing scheme based on semi-supervised learning, which jointly considersenergy efficiency, context-awareness, and optimal path configuration to enhance communication efficiency. Acontext-aware module is developed to collect and learn context information in an energy-efficient way and anew semi-supervised learning algorithm is proposed to estimate dynamic changes in network environment. Anovel routing metric is used to select the most reliable and stable path. Our simulation study shows that theproposed routing algorithm enhances the reliability and stability for CRSNs, and at the same time, significantlyimproves the packet delivery ratio.

1 INTRODUCTION

In the foreseeable future, tens of billions of electronicdevices will be expected to communicate with eachother and require a huge amount of radio resources.However, the current radio resources is lack of dueto inflexible spectrum sharing rules. In particular,existing wireless networks, such as WLANs, meshnetworks, body area networks, and sensor networks,which operate in unlicensed band will suffer from se-rious spectrum overcrowding problem. The cogni-tive radio technology that exploits dynamic spectrumsharing techniques is a promising solution to solvethis problem (Cesana et al., 2011).

Application of the cognitive radio technology inwireless sensor networks (WSNs) can open up newand unexplored network configuration possibilitiesand also enable researchers to explore new services.In Cognitive Radio Sensor Networks (CRSNs), thesensor devices are capable of sensing a wide spec-trum range, dynamically identifying available chan-nels, and intelligently accessing them. Unlike WSNs,CRSNs can operate in licensed bands. Cognitive Ra-dio Sensor Devices (CRSDs), which are also referredto as secondary users, share the licensed band withprimary users (PUs) who have higher priorities in oc-cupying the bands in a non-interfered manner. Thisindicates that the topology of CRSN is changing un-

predictably due to PUs’ activities and causes consid-erable difficulty in guaranteeing stable and efficientcommunications (Cesana et al., 2011; Akan et al.,2009; Ali et al., 2011).

There are many excellent prior research focusedon the lower layers (PHY/MAC) of cognitive radiotechnologies. However, routing, which is an im-portant requirement for efficient communication inmulti-hop based CRSNs, has not been well explored.Most of prior work on routing solutions are providedin ad-hoc based cognitive radio networks (Sampathet al., 2008; Cheng et al., 2007; Pefkianakis et al.,2008; Wang et al., 2009). These approaches employspectrum-aware schemes to support routing modulein the path selection process. However, these meth-ods do not take into account energy restriction of sen-sor networks, and thus cannot be directly applied toCRSN. There is only a limited work that focus onrouting issues in CRSNs. Parvin and Fujii (2011)and Shah et al. (2013) proposed spectrum-aware rout-ing solutions to guarantee network QoS requirements.In their proposed algorithms, the optimal path is se-lected to minimize end-to-end delays, but the impactof unpredictable link failure on QoS and communica-tion performance is not considered. In order to dy-namically predict the available spectrum resources inCRSN, a machine learning based routing algorithm isproposed in (Yu et al., 2010). The authors employ

188

Bayesian Learning to estimate the amount of avail-able resources, and then, the most reliable path thatcontains the largest number of available channels isselected. Unfortunately, this routing scheme is notfeasible because a large amount of labeled data, whichare used to train their learning algorithm, is difficultto obtain in CRSN.

This paper proposes a new energy efficient routingalgorithm that provides reliable communication per-formance in CRSN. One of the important features ofthe proposed routing scheme is that it can predict thePUs’ influence on spectrum usability and efficientlyevaluate link stability. In order to achieve this, acontext-aware module was developed to perceive con-text information such as PUs’ activity and varying ra-dio resources. A semi-supervised learning algorithmthat can provide good accuracy under limited labeleddata is developed to learn the context information, andthen predict the available radio sources in the future.Finally, a novel routing metric is defined to indicateend-to-end link stability, and a stable and reliable pathis selected based on a semi-Dijkstra algorithm.

The contribution of this paper can be summarizedas follows:

� Development of an energy-efficient context-awaremodule that integrates context information with acontext learning function.

� Development of a feasible learning method thatcan be applied to CRSNs where the available la-beled data are limited.

� Derivation of a novel routing metric to providestable and reliable paths.

The rest of the paper is organized as follows: Sec-tion 2 discusses the most relevant related work. Sec-tion 3 presents the proposed semi-supervised basedrouting algorithm. Section 4 validates the perfor-mance of our proposed algorithm through extensivesimulations. Finally, Section 5 concludes the paper.

2 RELATED WORK

Due to opportunistic channel access nature of the cog-nitive radio technology, routing techniques for tradi-tional WSNs are unable to satisfy the performance re-quirements of CRSNs.

There exist many routing techniques for ad-hocbased cognitive radio networks (CRNs). Sampathet al. (2008) and Cheng et al. (2007) proposedan AODV based spectrum-aware routing protocol forCRNs. Sampth et al. (2008) aim to guarantee end-to-end performance by integrating flow-based approachwith a link-based one. A routing metric is also derived

based on the number of available channels. Cheng,et al. (2007) propose an on-demand routing pro-tocol that selects suitable spectrums for each nodealong the path. According to spectrum availability,Pefkianakis et al. (2008) presented a routing solutioncalled SAMER to provide long-term and short-termroute, and Wang et al. (2009) proposed multi-pathbased routing protocol in order to improve connectionstability. These efforts mainly focused on route andspectrum selection and considered expected perfor-mance including throughput, delay, and robustness.However, these methods cannot be directly applied toCRSNs because the energy efficiency is not consid-ered as a design goal.

There is only limited work that focuses on routingissues in CRSNs. Parvin and Fujii (2011) proposed aspectrum-aware routing scheme with the goal of guar-anteeing network QoS. More specially, they define autility function which is used to evaluate the end-to-end delay, and a route with the maximum value is se-lected. Similar research is addressed by Shah et al.(2013). The authors design a distributed control algo-rithm to improve communication performance. Theirprimary goal is to guarantee the QoS requirements byoptimizing an objective function, which is derived tominimize queueing delay (Shah et al., 2013). How-ever, these routing schemes neglect the impact of un-predictable link failures on QoS.

In order to predict link failure and dynamicallychanging spectrum resources, a machine learningbased routing solution is proposed in (Yu et al., 2010).Yu et al. (2010) employ the Bayesian learning, whichis one of the supervised learning methods to estimatethe total number of neighboring PUs. The estimatedresult is used to reflect the amount of available ra-dio resources. Based on the estimation, the most reli-able path that contains the largest number of availablechannels is selected. The major shortcoming of thisrouting scheme is that the requirement of the super-vised learning based algorithm is ignored. In orderto make an accurate estimation, the Bayesian learningneeds a large amount of labeled data to train the learn-ing algorithms. Unfortunately, labeled data is hardto be obtained in most of CRSNs’ application sce-narios because the network environment dynamicallychanges. Furthermore, their route configuring schemeneglects dynamic link failures in CRSNs. Therefore,their routing scheme cannot guarantee reliable end-to-end communications.

In contrast to the aforementioned related work, theproposed routing algorithm predicts the PUs’ influ-ence and evaluates link stability to guarantee reliableand stable communications.

A�Routing�Algorithm�based�on�Semi-supervised�Learning�for�Cognitive�Radio�Sensor�Networks

189

3 PROPOSED SCHEME

The proposed routing algorithm consists of two parts:the Context-Aware module and Optimal Path Config-uration module. The responsibility of the context-aware module is to learn the characteristics of thevarying environment and predict the stability ofroutes. Based on that, the optimal path configurationmodule is used to estimate and select the most stablepath to guarantee end-to-end communication reliabil-ity. The following subsections discuss these parts indetail.

3.1 Context-aware Module

In order to feasibly estimate the variations in the net-work environment, semi-supervised learning, whichis a subcase of Machine Learning, is employed inour routing algorithm. Recently, Machine Learninghas become one of the most efficient and practicalsolutions to solve several routing issues (Yu et al.,2010; Wang et al., 2006; Ahmed and Kanhere, 2010).Supervised and unsupervised learning methods areparticular cases that perform learning tasks with la-beled and unlabeled data, respectively. Wang et al.(2006), and Ahmed and Kanhere (2010) applied thesupervised learning method to their research. How-ever, their algorithms require a large amount of la-beled data to train estimate functions. In most ofCRSNs application scenarios, the labeled data are dif-ficult and/or expensive to obtain. In contrast, theunsupervised learning method trains it’s estimationfunction based on unlabeled data, which are readilyavailable. However, the unsupervised method is morecomplex than its counterpart. Furthermore, the es-timation accuracy cannot satisfy the communicationrequirements of CRSNs. These shortcomings can bealleviated by semi-supervised learning, which bene-fits from tactfully utilizing both labeled data and un-labeled data. The semi-supervised learning methodcan provide good learning accuracy even when thereare only a few labeled data. Thus, it is more feasiblethan supervised or unsupervised method based one inCRSNs.

In order to learn the context information ofCRSNs, the routing algorithm needs to gather someinformation as labeled data. To predict the link con-nectivity to sink nodes, each node needs to maintainthe following context features:

� Neighbor node IDs (D1;D2; :::;Dn), and sink nodeIDs (S1;S2; :::;Sm).

� Current time slots (t1; t2; :::; tk) assuming that timeof day T is divided into k slots.

� Currently available channel set (Ch1;Ch2; :::;Chc)assuming a set of locally available channels.

In order to guarantee energy efficiency, the con-text information collection is performed in a passivemanner. In the network initialization step, the sinknode broadcasts HELLO message using a commoncontrol channel (CCC). In addition to the initializa-tion step, the message is broadcast repeatedly with aperiod (T + e)=k. During this period, e is randomlyselected within a time interval d (d� T=k) to pre-vent congestion. The HELLO message is one of thecontrol packets that contains sink nodes’ IDs, locallyavailable channel set, and a connectivity label Y (Yes)or N (No) field. A sink node sends the message af-ter setting the connectivity label as Y . This messageis identified by the sink node ID and the current timeslot. A CRSD may receive many copies of the samemessage from one of the neighbors. In this case,the CRSD only forwards the first one to downstreamneighbors.

When a node receives a HELLO message, itchecks its label and the available channel set of thesender. If the label is Y and they have the commonavailable channels, the node retains the label. Oth-erwise, the node changes the label to N. Before for-warding the message, the node stores the context in-formation, which consists of sink ID, sender ID, thecurrent time slot, and the label in local memory. Then,the localized context information is updated to themessage. The node stops the forwarding until thereare no downstream neighbors.

Through HELLO message forwarding, everyCRSD can maintain the labeled context informationand should know whether it is successfully connectedto a sink node. For example, a context informationhS1; t2;D5;fCh2;Ch3;Ch4g;Y i indicates that the nodecan successfully communicate with sink S1 at timeslot t2 using the relay node D5, and the available chan-nel set fCh2;Ch3;Ch4g.

The pseudo-code description of the proposed al-gorithm is given in Fig. 1. Let Li denote the labeledcontext information and Ui denote all possible com-bination of the unlabeled context information in Di.Two classifiers h1 and h2 are initially trained from Li.Then, the classifiers are used to label Ui � uk

i , whichindicate kth unlabeled context information in Di canbe labeled if both of the classifiers agree on the label-ing. When one of classifier disagree with the otherone, the decision is made after comparing with a con-fidence threshold value t. After labeling uk

i , the nodeadds it to labeled context information and then obtainsLnew

i = Li[uki . Then, the algorithm picks up uk+1

i andrepeats this procedure. This process is repeated un-til there are no more confident unlabeled data to be

SENSORNETS�2014�-�International�Conference�on�Sensor�Networks

190

Figure 1: Pseudo-code for the learning algorithm.

selected.In this paper, we employ the naive-Bayes classi-

fier and decision tree learning as the classifiers h1 andh2, respectively. The naive-Bayes classifier, which isbased on the Bayes rule, is widely employed in pos-terior probability calculation with priori information.The naive-Bayes classifier is defined by the followingequation:

h1(uki ) = argmax

l2fY;NgP(l)Õ

jP(X jjl); (1)

where X j represents the context information value andl denotes the labels (l 2 fY;Ng).

The decision tree learning makes decision by di-viding the classification into a set of choices, andstarting at the root of the tree and progressing downto the leaves (Marsland, 2009). Usually, the attributethat has the highest information gain for the label isselected as the root node and the parent nodes. Theinformation gain can be calculated using the entropyof the attribute (Marsland, 2009).

The following example illustrates the proposed al-gorithm. Fig. 2 shows a CRSN where CRSDs oc-cupy four licensed channels. We assume that a PUis moving through the CRSN during t1 to t2. Withthe HELLO message traverses the entire network, D2collects the localized context information as shown inTable 1, which shows both the labeled and unlabeledcontext information. After training the classifiers withthe labeled data, they are used to classify the unla-beled data (the last four rows of the table). When thealgorithm selects an unlabeled data hS; t1;D4;Ch4i,which is obtained form the 5th row in the table, theinference results is shown as follows:

P(Y ) Õj=4

P(X j=4jY ) =P(Y )P(SjY )P(t1jY )P(D1jY )P(Ch4jY )

= 810�1� 3

8�58�

18

=0:023;

(2)

Table 1: Labeled and unlabeled data in the CRSD.

DestinationTimeslot

NeighborAvailablechannels

Label

S t1 D4 Ch1;Ch2;Ch3 YS t1 D5 Ch2;Ch4 NS t2 D4 Ch1;Ch3 YS t2 D5 Ch2;Ch3;Ch4 YS t1 D4 Ch4 ?(N)

S t2 D4 Ch2 ?(Y )S t2 D4 Ch4 ?(Y )S t2 D5 Ch1 ?(Y )

Figure 2: Context information gathering.

P(N) Õj=4

P(X j=4jN) =P(N)P(SjN)P(t1jN)P(D1jN)P(Ch4jN)

= 14�1�1�

0+mpD21+m �1

=0:042; where m=10; pD2=12 :

(3)

Eq. (2), and Eq. (3) indicate the packet de-livery probability in Y and N cases, respectively.By substituting the results to Eq. 1, there isP(N) Õ

j=4P(X j=4jN) > P(Y ) Õ

j=4P(X j=4jY ), and then

the algorithm obtains the inference result as:

h1(u12) = N: (4)

The other inference result can be obtained afterconstructing a decision tree1 as shown in Fig. 3.Based on the decision tree, the inference decision be-comes:

h2(u12) = N: (5)

Note that the intersection of the inference results isnot empty; therefore, the unlabeled data is labeled ash1(u1

2)\h2(u12) = N. Then, the labeled data set L2 is

updated with the labeled data u12 = hS; t1; D4;Ch4;Ni,

and a new set of labeled data is Lnew2 = L2 [ u1

2. Theaugmented labeled set can then be used to retrain theclassifiers in an iterative manner until the terminationcondition is satisfied.

1We skip derivation process of the decision tree. Read-ers can refer to Chapter 6 in (Marsland, 2009).


191

Figure 3: The decision tree.

Finally, the labeled context information is storedin local memory to further predict the stability ofroutes, and also is used to derive a novel routing met-ric for selecting optimal path as discussed in the fol-lowing subsection.

3.2 Optimal Path Configuration

In order to evaluate the benefit of an intermediatenode for path stability, a new routing metric calledPath Stable Metric (PSM) is defined as follows:

PSM = q åj2c

h1(unewi )+(1�q)

1nc

åj2c

h2(unewi ); (6)

where c and nc denote the available channel set andthe number of channels in c, respectively, and q (0�q� 1) is a weighted parameter for controlling the pro-portion of the two classifiers.

In this paper, we assume that only one channel canbe occupied at a time. This indicates that any chan-nel in the available channel set is independent of theothers. Therefore, the routing metric must be derivedby cumulating each channel’s stability. The first termin the right side of the equation is obtained by h1(�).The other one is obtained from h2(�). If the label is Y ,h2(�) = 1; otherwise, h2(�) = 0.

When a CRSD receives a routing request fromthe upper layer, it broadcasts route request (RREQ)packets (Perkins and Royer, 1999) on CCC. The mes-sage contains sink node ID and its currently avail-able channel set. If the neighbors who receive thismessage have the common available channels withthe sender, the nodes calculate local PSMs. Beforeforwarding the RREQ, intermediate nodes add its IDand PSM, and update available channel informationwith locally available channels. Finally, when the sinknode receives the RREQ, a connectivity diagram isconstructed based on contained intermediate nodes’

Figure 4: Optimal path selection.

ID and PSM information. Note that the sink node canreceive multiple copies of the same RREQ packetsfrom different neighbor nodes. In that case, the con-nectivity diagram should be updated only if a mes-sage contains a new PSM information. Every up-dated topology will trigger the semi-Dijkstra algo-rithm to find the path with the highest PSM. Then, thesink node sends a route reply (RREP) packet throughthe selected path. The RREP packet contains all thenodes’ IDs along the path; therefore, the source nodecan send data packets to the sink node along the moststable path.

In order to illustrate the optimal path configura-tion, consider again the context data learning exam-ple discussed in Subsection 3.1 and shown in Fig.4. Suppose that node D1 broadcasts a RREQ mes-sage, which consists of the current context informa-tion hS; t2;fCh1;Ch2gi, at time slot t2. When theneighbor node D2 receives the message, it checks theintersection of available channel set. If they havea common set of available channels (In this case,their common available channel set is fCh1;Ch2g \fCh1;Ch2;Ch3;Ch4g = fCh1;Ch2g, and q = 0:7 ),then the PSM is calculated as follows:

q åj2c

h1(unew1 ) = q(P(Y )P(SjY )P(t2jY )P(Ch1jY )

+P(Y )P(SjY )P(t2jY )P(Ch2jY ))= 0:182;

(7)

(1�q)1nc

åj2c

h2(unew1 ) = (1�0:7)

12(1+1) = 0:3:

(8)By substituting the results to Eq. (9), PSM for D2 isgiven by

PSMD2 = q åj2c

h1(unew1 )+(1�q)

1nc

åj2c

h2(unew1 ) = 0:482:

(9)PSMs for the other nodes in Fig. 4 are derived in

the same manner.After receiving the RREQ message, the sink node

can construct a network topology as shown in Fig.


192

4. Using the semi-Dijkstra’s algorithm, the sink nodecan find the most reliable and stable path and then re-ply to D1 with a RREP.

4 PERFORMANCE EVALUATION

In this section, we evaluate the performance of theproposed scheme through simulation. We randomlydeploy 100 CRSDs in a simulated area 100� 50 m2.The packet size is set to 100 bytes, and one CCC andfour licensed channels are available in the simulationenvironment. We assume that the transmission rangeof CRSD is 5 m and the interference rage is 10 m. Theactivity of PUs is a Poisson process with arrival rate l.We also assume that CRSD stops transmission imme-diately if it suffers from PUs’ interference. Since thenaive-Bayes classifier has been shown to be more sta-ble and provide better performance than decision treebased methods (Marsland, 2009), the proposed rout-ing algorithm uses q=0:7 indicating that more trust isgiven to the naive-Bayes classifier.

The proposed routing algorithm is compared withSAMER (Pefkianakis et al., 2008), which uses theminimum hop count as the route selection metric. Fig.5 compares the communication reliability of the tworouting algorithms. For each simulation run, the ar-rival rate l is randomly selected in the interval of0:15�0:5. We can observe from Fig. 5 that the packetdelivery ratio is not stable when the network employsminimum hop-count as the routing metric. This is be-cause the performance of SAMER is significantly af-fected by how often the minimum hop-count path isinterrupted by PUs. As the simulation proceeds, theaverage packet delivery ratio of SAMER convergesto 72% which cannot satisfy most QoS requirements.In comparison, the proposed semi-supervised basedrouting algorithm shows more stable packet deliveryratio and the average performance converged to 92%.More specifically, the proposed routing protocol pro-vides 28% performance gain over SAMER. There aretwo main reasons for this. First, the proposed ContextAware module can intelligently estimate the scalabil-ity of the path. Second, the most stable path, whichhas the lowest possibility of interruption, can be ef-fectively established based on the proposed routingmetric. Furthermore, the proposed method can effi-ciently avoid PUs’ influence; therefore, extra energyconsumption caused by frequent retransmissions canalso be efficiently avoided.

Fig. 6 shows that the proposed method clearlyoutperforms SAMER in terms of throughput. Asthe arrival rate of PUs increases, the throughput forSAMER drastically decreases. This is because the ac-

Figure 5: Packet delivery ratio.

Figure 6: The network throughput

tive behavior of PUs has a significant impact on thesystem throughput. Although SAMER has benefits interms of minimum hop routing and low transmissiondelay, it cannot provide satisfactory throughput dueto the influence of PUs. The proposed routing algo-rithm shows more stable performance than SAMER,even when the arrival rate of PUs is increased. This isdue to the fact that the proposed routing algorithm re-duces the probability of route failure by learning andestimating the activity and channel utilization of PUs.

5 CONCLUSIONS

This paper presented a semi-supervised learningbased routing algorithm that predicts and minimizesthe influence of PUs. The proposed routing algorithmconsists of Context-Aware and Optimal Path Config-uration modules, and uses a feasible semi-supervisedlearning algorithm to perceive variation of the net-


193

work connectivity. In addition, a novel routing metricis derived to estimate end-to-end connection stabil-ity. The optimal path, which is the most stable in thenetwork topology, is selected using the semi-Dijkstraalgorithm. Finally, our simulation study shows thatthe proposed method effectively provides reliable andstable paths, and improves system throughput. Weare currently in the progress of designing an energy-efficient cross-layer routing algorithm, which com-bines spectrum selection with route configuration byapplying a feasible machine learning method.

ACKNOWLEDGEMENTS

This work was supported by the MSIP(Ministry ofScience, ICT & Future Planning), Korea, under theITRC(Information Technology Research Center) sup-port program supervised by the NIPA(National ITIndustry Promotion Agency) (NIPA-2013-(H0301-13-2001)) and by Mid-career Researcher Programthrough NRF grant funded by the MEST (No. 2011-0015744).

REFERENCES

Ahmed, S. and Kanhere, S. (2010). A bayesian routingframework for delay tolerant networks. In Proc. ofWCNC’10.

Akan, O. B., Karli, O. B., and Ergul, O. (2009). Cogni-tive radio sensor networks. IEEE Network Magazine,23(4).

Ali, A., Iqbal, M., Baig, A., and Wang, X. (2011). Rout-ing techniques in cognitive radio networks: A survey.International Journal of Wireless & Mobile Networks,3(3).

Cesana, M., Cuomo, F., and Ekici, E. (2011). Routing incognitive radio network: Challenges and solutions. In-ternational Journal on Ad Hoc Networks, 9(3).

Cheng, G., Liu, W., and Cheng, W. (2007). Spectrum awareon-demand routing in cognitive radio networks. InProc. of DySPAN’07.

Marsland, S. (2009). Machine Learning: An AlgorithmicPerspective. A Chapman & Hall.

Pefkianakis, I., Wong, S. H. Y., and Lu, S. (2008). Samer:Spectrum aware mesh routing in cognitive radio net-works. In Proc. of DySPAN’08.

Perkins, C. E. and Royer, E. M. (1999). Ad-hoc on-demanddistance vector routing. In Proc. of WMCSA’99.

Sampath, A., Yang, L., Cao, L., and Zhao, B. Y. (2008).High throughput spectrum-aware routing for cognitiveradio networks. In Proc. of CROWNCOM’08.

Shah, G., Gungor, V., and Akan, O. (2013). Routing in cog-nitive radio network: Challenges and solutions. IEEETran. on Industrial Informatics.

Wang, X., Kwon, T., and Choi, Y. (2009). A multipath rout-ing and spectrum access framework for cognitive ra-dio systems in multi-radio networks. In Proc. of ACMWorkshop on Cognitive Radio Networks.

Wang, Y., Martonosi, M., and Peh, L. S. (2006). Super-vised learning in sensor networks: new approacheswith routing, reliability optimizations. In Proc. ofSECON’06.

Yu, R., Zhang, Y., Yao, W. Q., Song, L. Y., and Xie, S. L.(2010). Spectrum-aqare routing for reliable end-to-end communications in cognitive sensor networks. InProc. of IEEE Globecom’10.


194

A Routing Algorithm based on Semi-supervised Learning for Cognitive Radio...

Documents

Transcript of A Routing Algorithm based on Semi-supervised Learning for Cognitive Radio...