Performance analysis of machine learning algorithms on self localization system1
Download as DOCX, PDF0 likes13 views
The document discusses various machine learning algorithms for performance analysis in self-localization systems, including SVM, decision trees, Naïve Bayes, random forests, k-nearest neighbors, bagging classifiers, AdaBoost, and multilayer perceptrons. It highlights that the random forest algorithm generally provides the best prediction accuracy among these methods, while also detailing their individual functionalities and applications. Python was used to implement these algorithms, utilizing a 'student data' dataset along with several libraries for data processing and model evaluation.
![Venkat Java Projects
Mobile:+91 9966499110
Visit:www.venkatjavaprojects.com Email:venkatjavaprojects@gmail.com
K-NN isa type of instant-basedlearning,orlazylearning, where the functionisonly
approximatedlocallyandall computationisdeferreduntil classification.
Both forclassificationandregression,auseful technique canbe toassignweightstothe
contributionsof the neighbors,sothatthe nearerneighborscontribute more tothe average
than the more distantones.Forexample,acommonweightingscheme consistsingiving
each neighboraweightof 1/d,where d isthe distance tothe neighbor
The neighborsare takenfroma setof objectsforwhichthe class (for k-NN classification)or
the objectpropertyvalue (for k-NN regression) isknown.Thiscanbe thoughtof as the
trainingsetforthe algorithm,thoughnoexplicittrainingstepisrequired.
A peculiarityof the k-NN algorithmisthatitis sensitive tothe local structure of the data.
Bagging classifier:
A Baggingclassifierisanensemble meta-estimatorthatfitsbase classifierseachonrandomsubsets
of the original datasetandthenaggregate theirindividual predictions(eitherbyvotingorby
averaging) toforma final prediction.Suchameta-estimatorcantypicallybe usedasa wayto reduce
the variance of a black-box estimator(e.g.,adecisiontree),byintroducingrandomizationintoits
constructionprocedure andthenmakinganensemble outof it.
Each base classifieristrainedinparallelwithatrainingsetwhichisgeneratedbyrandomlydrawing,
withreplacement,N examples (ordata) fromthe original trainingdataset –whereN is the size of the
original training set.Trainingsetfor each of the base classifiersisindependentof eachother.Many
of the original datamaybe repeatedinthe resultingtrainingsetwhile othersmaybe leftout.
Baggingreducesoverfitting(variance) byaveragingorvoting,however,thisleadstoanincrease in
bias,whichis compensatedbythe reductioninvariance though.
AdaBoost:
Adaptiveboosting isamachine learningmeatalgorithm formulated.Itcanbe usedin conjunction
withmanyothertypesof learningalgorithmstoimprove performance.The outputof the other
learningalgorithms('weaklearners') iscombinedintoaweightedsumthatrepresentsthe final
outputof the boostedclassifier.AdaBoostisadaptive inthe sense thatsubsequentweaklearners
are tweakedinfavorof those instancesmisclassifiedbypreviousclassifiers.AdaBoostissensitiveto
noisydata and outliers. Insome problemsitcanbe lesssusceptible tothe overfittingproblemthan
otherlearningalgorithms.The individuallearnerscanbe weak,butas longas the performance of
each one isslightlybetterthanrandomguessing,the final model canbe proventoconverge toa
stronglearner.
Everylearningalgorithmtendsto suitsome problemtypesbetterthanothers,andtypicallyhas
manydifferentparametersandconfigurationstoadjustbefore itachievesoptimalperformance ona
dataset,AdaBoost isoftenreferredtoasthe bestout-of-the-box classifier.[2]
Whenusedwith
decisiontree learning,informationgatheredateachstage of the AdaBoostalgorithmaboutthe
relative 'hardness'of eachtrainingsample isfedintothe tree growingalgorithmsuchthatlatertrees
tendto focuson harder-to-classifyexamples.](https://image.slidesharecdn.com/performanceanalysisofmachinelearningalgorithmsonselflocalizationsystem1-200318103328/85/Performance-analysis-of-machine-learning-algorithms-on-self-localization-system1-3-320.jpg)
Performance analysis of machine learning algorithms on self localization system1
- 1. Venkat Java Projects Mobile:+91 9966499110 Visit:www.venkatjavaprojects.com Email:[email protected] Performance Analysis of Machine Learning Algorithms on Self Localization Systems In thispaperauthoris usingSVM(SupportVectorMachine), DecisionTree Classifier,K-Neighbors Classifier,naïve bayes,RandomForestClassifier,BaggingClassifier,AdaBoostClassifierandMLP Classifier All the algorithmsgenerate model fromtraindatasetandnew datawill be appliedontrainmodel to predictitclass.Random Forest algorithmisgivingbetterpredictionaccuracycompare to all other algorithm. Support vector machine: Machine learninginvolvespredictingandclassifyingdataandto do sowe employvariousmachine learningalgorithmsaccordingtothe dataset.SVMor SupportVectorMachine is a linearmodel for classificationandregressionproblems.Itcansolve linearandnon-linearproblemsandworkwell for manypractical problems.The ideaof SVMissimple:The algorithmcreatesaline or a hyperplane whichseparatesthe dataintoclasses.Inmachine learning,the radial basisfunctionkernel,orRBF kernel,isapopularkernel functionusedinvariouskernelizedlearningalgorithms.Inparticular,itis commonlyusedinsupportvectormachine classification.Asasimple example,foraclassification task withonlytwofeatures(like the image above),youcanthinkof a hyper plane asa line that linearlyseparatesandclassifiesasetof data. Intuitively,the furtherfromthe hyper plane ourdatapointslie,the more confidentwe are thatthey have beencorrectlyclassified.We therefore wantourdatapointstobe as far awayfrom the hyper plane as possible,while still beingonthe correctside of it. So whennewtestingdataisadded, whateversideof the hyperplane itlandswill decide the class that we assignto it. How dowe findthe righthyper plane? Or, inotherwords,howdo we bestsegregate the twoclasseswithinthe data? The distance betweenthe hyper plane andthe nearestdatapointfromeithersetisknownasthe margin.The goal isto choose a hyper plane withthe greatestpossible marginbetweenthe hyper plane andany pointwithinthe trainingset,givingagreaterchance of new data beingclassified correctly. Both algorithmsgeneratemodel fromtraindatasetandnew datawill be appliedontrainmodel to predictitclass.SVMalgorithmisgivingbetterpredictionaccuracycompare to ANN algorithm. Naïve Bayes Classifier Algorithm It wouldbe difficultand practicallyimpossible toclassifyawebpage,a document,anemail orany otherlengthytextnotesmanually.This iswhere Naïve BayesClassifiermachine learningalgorithm comesto the rescue.A classifierisafunctionthatallocatesa population’s elementvalue fromone of the available categories.Forinstance,SpamFilteringisapopularapplicationof Naïve Bayes algorithm.Spamfilterhere,isaclassifierthatassignsalabel “Spam”or “Not Spam” to all the emails.
- 2. Venkat Java Projects Mobile:+91 9966499110 Visit:www.venkatjavaprojects.com Email:[email protected] Naïve BayesClassifierisamongst the mostpopularlearningmethodgroupedbysimilaritiesthat workson the popularBayesTheorem of Probability- tobuildmachine learningmodelsparticularly for disease predictionanddocumentclassification.Itisa simple classification of wordsbased on BayesProbabilityTheoremforsubjective analysisof content. Decision tree: A decisiontree isagraphical representationthatmakesuse of branchingmethodologytoexemplify all possible outcomesof adecision, basedoncertainconditions.Inadecisiontree,the internalnode representsateston the attribute,eachbranchof the tree represents the outcome of the testand the leaf node representsaparticularclasslabel i.e.the decisionmade aftercomputingall of the attributes. The classificationrulesare representedthroughthe pathfromrootto the leaf node. Types ofDecisionTrees ClassificationTrees- These are consideredasthe defaultkindof decisiontreesusedtoseparate a datasetintodifferentclasses,based onthe responsevariable.Theseare generallyusedwhenthe response variable iscategoricalinnature. RegressionTrees-Whenthe responseortargetvariable iscontinuousornumerical,regressiontrees are used.These are generally usedinpredictive type of problemswhencomparedtoclassification. Decisiontreescanalsobe classifiedintotwotypes,basedonthe type of targetvariable- Continuous Variable DecisionTreesandBinary Variable DecisionTrees.Itisthe targetvariable thathelpsdecide whatkindof decisiontree wouldbe requiredforaparticularproblem. Random forest: RandomForestisthe go to machine learningalgorithmthatusesabaggingapproachto create a bunchof decisiontreeswithrandom subsetof the data.A model istrainedseveral timesonrandom sample of the datasetto achieve goodpredictionperformancefromthe randomforestalgorithm.In thisensemble learningmethod,the outputof all the decisiontreesinthe randomforest,is combinedtomake the final prediction.The final prediction of the randomforestalgorithmisderived by pollingthe resultsof eachdecisiontree orjustby goingwitha predictionthatappearsthe most timesinthe decisiontrees. For instance,inthe above example- if 5 friendsdecide thatyouwill like restaurantRbutonly2 friendsdecide thatyouwill notlike the restaurantthenthe final predictionisthat,youwill like restaurantR as majorityalwayswins. K – nearest neighbor: K-nearest neighbor’salgorithm(k-NN) isa nonparametricmethodused forclassification and regression Inbothcases,the inputconsistsof the kclosesttrainingexamplesinthe feature space.The outputdependsonwhether k-NN isusedforclassificationorregression: In k-NN classification,the outputisa class membership.Anobjectisclassifiedbya pluralityvote of itsneighbors,withthe objectbeingassignedtothe classmost common amongits k nearestneighbors(kisapositive integer,typicallysmall).If k= 1, thenthe objectissimplyassignedtothe classof that single nearestneighbor. In k-NN regression,the outputisthe propertyvalue forthe object.Thisvalue isthe average of the valuesof k nearestneighbors.
- 3. Venkat Java Projects Mobile:+91 9966499110 Visit:www.venkatjavaprojects.com Email:[email protected] K-NN isa type of instant-basedlearning,orlazylearning, where the functionisonly approximatedlocallyandall computationisdeferreduntil classification. Both forclassificationandregression,auseful technique canbe toassignweightstothe contributionsof the neighbors,sothatthe nearerneighborscontribute more tothe average than the more distantones.Forexample,acommonweightingscheme consistsingiving each neighboraweightof 1/d,where d isthe distance tothe neighbor The neighborsare takenfroma setof objectsforwhichthe class (for k-NN classification)or the objectpropertyvalue (for k-NN regression) isknown.Thiscanbe thoughtof as the trainingsetforthe algorithm,thoughnoexplicittrainingstepisrequired. A peculiarityof the k-NN algorithmisthatitis sensitive tothe local structure of the data. Bagging classifier: A Baggingclassifierisanensemble meta-estimatorthatfitsbase classifierseachonrandomsubsets of the original datasetandthenaggregate theirindividual predictions(eitherbyvotingorby averaging) toforma final prediction.Suchameta-estimatorcantypicallybe usedasa wayto reduce the variance of a black-box estimator(e.g.,adecisiontree),byintroducingrandomizationintoits constructionprocedure andthenmakinganensemble outof it. Each base classifieristrainedinparallelwithatrainingsetwhichisgeneratedbyrandomlydrawing, withreplacement,N examples (ordata) fromthe original trainingdataset –whereN is the size of the original training set.Trainingsetfor each of the base classifiersisindependentof eachother.Many of the original datamaybe repeatedinthe resultingtrainingsetwhile othersmaybe leftout. Baggingreducesoverfitting(variance) byaveragingorvoting,however,thisleadstoanincrease in bias,whichis compensatedbythe reductioninvariance though. AdaBoost: Adaptiveboosting isamachine learningmeatalgorithm formulated.Itcanbe usedin conjunction withmanyothertypesof learningalgorithmstoimprove performance.The outputof the other learningalgorithms('weaklearners') iscombinedintoaweightedsumthatrepresentsthe final outputof the boostedclassifier.AdaBoostisadaptive inthe sense thatsubsequentweaklearners are tweakedinfavorof those instancesmisclassifiedbypreviousclassifiers.AdaBoostissensitiveto noisydata and outliers. Insome problemsitcanbe lesssusceptible tothe overfittingproblemthan otherlearningalgorithms.The individuallearnerscanbe weak,butas longas the performance of each one isslightlybetterthanrandomguessing,the final model canbe proventoconverge toa stronglearner. Everylearningalgorithmtendsto suitsome problemtypesbetterthanothers,andtypicallyhas manydifferentparametersandconfigurationstoadjustbefore itachievesoptimalperformance ona dataset,AdaBoost isoftenreferredtoasthe bestout-of-the-box classifier.[2] Whenusedwith decisiontree learning,informationgatheredateachstage of the AdaBoostalgorithmaboutthe relative 'hardness'of eachtrainingsample isfedintothe tree growingalgorithmsuchthatlatertrees tendto focuson harder-to-classifyexamples.
- 4. Venkat Java Projects Mobile:+91 9966499110 Visit:www.venkatjavaprojects.com Email:[email protected] Multilayer perceptron (MLP): A multilayerperceptron(MLP) isa classof feedforward artificialneural network (ANN).The term MLP isusedambiguously,sometimeslooselytoreferto any feed forwardANN,sometimesstrictlyto referto networkscomposedof multiple layersof perceptrons (withthresholdactivation); see § Terminology.Multilayerperceptronsare sometimescolloquiallyreferredtoas"vanilla"neural networks,especiallywhentheyhave asingle hiddenlayer. An MLP consistsof at leastthree layersof nodes:aninputlayer,a hiddenlayerandanoutputlayer. Exceptfor the inputnodes,eachnode isa neuronthatusesa nonlinear activationfunction.MLP utilizesasupervised learningtechnique called backpropagation fortraining. Itsmultiple layersand non-linearactivationdistinguishMLPfroma linear perceptron.Itcandistinguishdatathatis not linearlyseparable. To implementabove all algorithmswe have usedpython technologyand‘student data’ dataset.This dataset available inside dataset folder which contains test dataset with dataset information file. PythonPackagesandLibrariesused:Numpy,pandas, tkinter, PyVISA 1.10.1 1.10.1 PyVISA-py 0.3.1 0.3.1 cycler 0.10.0 0.10.0 imutils 0.5.3 0.5.3 joblib 0.14.1 0.14.1 kiwisolver 1.1.0 1.1.0 matplotlib 3.1.2 3.1.2 nltk 3.4.5 3.4.5 numpy 1.18.1 1.18.1 opencv-python 4.1.2.30 4.1.2.30 pandas 0.25.3 0.25.3 pip 19.0.3 20.0.1 pylab 0.0.2 0.0.2 pyparsing 2.4.6 2.4.6 python-dateutil 2.8.1 2.8.1 pytz 2019.3 2019.3 pyusb 1.0.2 1.0.2 scikit-learn 0.22.1 0.22.1 scipy 1.4.1 1.4.1 seaborn 0.9.0 0.9.0 setuptools 40.8.0 45.1.0 six 1.14.0 1.14.0 sklearn 0.0 0.0 style 1.1.6 1.1.6
- 5. Venkat Java Projects Mobile:+91 9966499110 Visit:www.venkatjavaprojects.com Email:[email protected] styled 0.2.0.post1 0.2.0.post1 classificationreport,confusionmatrix,accuracyscore,train_test_split,K-Fold,cross_val_score,Grid SearchCV, DecisionTree Classifier,K-NeighborsClassifier,SVC,naive_bayes,Random Forest Classifier,BaggingClassifier,AdaBoostClassifier,MLP Classifier. Screenshots Whenwe run the code itdisplaysbelow window Nowclickon ‘uploaddataset’touploadthe data
- 6. Venkat Java Projects Mobile:+91 9966499110 Visit:www.venkatjavaprojects.com Email:[email protected] Nowclickon ‘readdata’ itreads the data Nowclickon ‘Train_Test_split’tosplitthe dataintotrainingandtesting
- 7. Venkat Java Projects Mobile:+91 9966499110 Visit:www.venkatjavaprojects.com Email:[email protected] Nowclickon ‘All classifiers’toclassifythe models KNN Predicted Values on Test Data is 98% CART Predicted Values on Test Data is 97.31% SVM Predicted Values on Test Data is 98.18% RF Predicted Values on Test Data is 98.62% Bagging Predicted Values on Test Data is 97.41% Ada Predicted Values on Test Data is 87.43% MLP Predicted Values on Test Data is 98.00% Nowclickon ‘Model comparison’the comparisonbetweenthe models
- 8. Venkat Java Projects Mobile:+91 9966499110 Visit:www.venkatjavaprojects.com Email:[email protected]