A Survey on Machine Learning Algorithms for Applications in Cognitive Radio Networks

Cognitive Radio with Machine Learning to Increase Spectral Efficiency in Indoor Applications on the 2.5 GHz Band

“…Overall, the SVM and RF algorithms have a higher accuracy (i.e., lower ∆), typically within 5% of the real value—the two classifiers being less sensitive to changes in occupancy rates and operator traffic. These graphs confirm the research presented in [ 17 ], in which it was reported that spectrum sensing is a task well-performed by algorithms such as kNN, Q-learning, RF, and SVM.…”

“…Figure 14 also reinforces the use of RF for in loco detection of the occupancy rate, as it presents the best accuracy when compared to SVM in datasets with 10% of the RBs of each incumbent—keeping ∆ typically within 5% of the real value—mainly in real occupancy rates below 30%, where the precision of the in loco evaluation must be high. These graphs confirm the research presented in [ 17 ], in which it was reported that spectrum sensing is a task well-performed by algorithms such as kNN, Q-learning, RF, and SVM.…”

“…For the past several years, it has been possible to find examples of works that attempt to address how a CR can operate opportunistically in channels used by different PUs, where spectrum sensing is an essential component. CR, DSA, and ML algorithms are powerful techniques for designing a promising spectrum sensing model to improve IoT on the licensed spectrum [ 2 , 3 , 6 , 9 , 17 ]. For example, in [ 5 ] the authors used the licensed spectrum with the supervised ML algorithms support vector machine (SVM), k-nearest neighbor (kNN), and decision tree (DT) to detect the existence of primary users (PUs) over the television band.…”

“…All of the experiments reported here were performed in Python with Sklearn and standard parameters, where each tested dataset used 30% of the data for training. After checking the technical literature about CR for spectrum sensing and spectrum prediction [ 5 , 17 , 28 , 37 , 38 , 39 , 40 , 41 , 42 ], we obtained datasets with different occupancy rates by changing the RSRP decision threshold with 0.5 dB steps and the information of all of the 50 or 100 RBs. Thus, we evaluated the following classifiers: naïve Bayes (NB), random forest (RF) , multilayer perceptron neural network (MP), and support vector machine (SVM) algorithms.…”

“…Thus, we evaluated the following classifiers: naïve Bayes (NB), random forest (RF) , multilayer perceptron neural network (MP), and support vector machine (SVM) algorithms. These classifiers were selected to cover a wide range of algorithm families [ 5 , 17 , 37 , 38 , 39 , 40 , 41 , 42 ].…”