Metode razvoja i adaptacije regresionih modela bazirane na genetskim algoritmima

Abstract

Većina postojećih regresionih metoda modeliranja pretpostavlja vremensku nepromenljivost modeliranih objekata i zahteva stalan skup ulaznih parametara. U realnim aplikacijama, stalne promene objekata i otkazi merne opreme mogu dovesti do situacija u kojima usvojeni model postaje neupotrebljiv. Iz tog razloga je neophodno razviti metode i sisteme za automatsko generisanje što adekvatnijih modela za datu situaciju. U okviru ove disertacije su razvijena dva hibridna metoda koji nude deo rešenja za navedene probleme. MLR/GA hibrid omogućava generisanje linearnog regresionog modela (MLR) koji je, za date uslove, optimizovan pomoću genetskih algoritama po kriterijumu tačnosti i kriterijumu kompleksnosti. Za razliku od postojećih metoda, MLR/GA metod omogućava generisanje adaptivnih modela koji su otporni na promenljivost skupa ulaznih promenljivih i promenljivost skupa izmerenih vrednosti. Razvijeni MLR/GA metod je implementiran u vidu GenReg softverskog agenta, čije performanse su testirane u postupku modeliranja radijalnog pomeranja odabranih tačaka betonske brane Bočac, na reci Vrbas u Republici Srpskoj. Modeli generisani korišćenjem MLR/GA metoda su u slučaju otkaza pojedinih senzora pokazali značajno bolju sposobnost za predikciju u odnosu na MLR modele koji podrazumevaju stalan skup ulaznih promenljivih. Dodatno, hibridni metod je pokazao sposobnost da pri generisanju modela odbacuje prediktore koji nisu od značaja za opisivanje posmatranog objekta. ANN/GA je hibridni metod za razvoj i adaptaciju regresionih modela zasnovanih na veštačkim neuronskim mrežama (ANN). Korišćenjem genetskih algoritama ANN/GA metod optimizuje strukturu i parametre neuronske mreže u skladu sa aktuelnim skupovima ulaznih i izlaznih promenljivih, i merenih vrednosti. Za razliku od sličnih postojećih rešenja, ANN/GA metod optimizuje skoro sve elemente neuronske mreže. Hibrid vrši samopodešavanje modela tako što optimizuje broj skrivenih slojeva, broj neurona u tim slojevima, izbor aktivacione funkcije, algoritam učenja, kao i vrednosti parametara učenja u skladu sa odabranim algoritmom. Razvijeni ANN/GA metod je implementiran u vidu DEVONNA softverskog agenta koji je validovan kroz studiju slučaja brane Grančarevo, na reci Tebišnjici u Republici Srpskoj, a rezultati su poređeni sa rezultatima dobijenim korišćenjem ekvivalentnog MLR/GA hibrida. Realizovani testovi su pokazali da modeli generisani ANN/GA hibridom mogu dati predikcije strukturnog ponašanja brane sa većom tačnošću od MLR modela. Međutim, za razliku od modela u obliku MLR, koji su otporni na temperaturne fazne pomake prisutne na različitim geografskim lokacijama, modeli u formi ANN pokazuju nestabilno ponašanje pod takvim okolnostima. Pored toga, generisanje ANN modela je vremenski znatno zahtevnije. Komparativna analiza modela generisanih na osnovu MLR/GA i ANN/GA metoda sa jedne, i modela u formi postepenih regresija, sa druge strane, je pokazala da predstavljeni metodi u pojedinim aspektima prevazilaze mogućnosti postojećih metoda za generisanje regresionih modela. Uz primenu tehnika redukcije dimenzija prostora istraživanja, predloženi hibridni metodi i razvijeni softverski agenti predstavljaju moćan alat za modeliranje realnih objekata i sistema.

Most of existing regression modeling methods presuppose the time immutability of the modeled objects and require a constant set of input parameters. In real applications, the constant changes of the objects and failures of measuring equipment can lead to situations in which the adopted model becomes unusable. For this reason it is necessary to develop methods and systems for automatic generation of the most adequate models for the given situation. In this dissertation two hybrid methods that offer part of the solution to the above problems have been developed. MLR/GA hybrid is able to generate a linear regression model (MLR) which is, for the given conditions, optimized by using genetic algorithms according to the criterion of accuracy and complexity criterion. Unlike the existing methods, MLR/GA method is enable to generate the adaptive models that are resistant to the variability of the set of input variables and the growing set of measured values. The developed MLR/GA method is implemented in the form of GenReg software agent, whose performances have been tested in the process of modeling the radial displacement of the selected points of Bočac concrete dam on the Vrbas river, in the Republic of Srpska. In the case of failure of individual sensors, models generated by using MLR/GA method showed a significantly better prediction compared to the MLR models that implied a constant set of input variables. In addition, the hybrid method has shown the capability of rejecting predictors that have no influence on the modeled object. ANN/GA is a hybrid method for the development and adaptation of regression models based on artificial neural networks (ANN). Using genetic algorithms ANN/GA method optimizes the structure and parameters of neural network in accordance with the current sets of input and output variables and measured values. Unlike similar existing solutions, ANN/GA method optimizes nearly all the elements of a neural network. The hybrid performs self-tuning of the model by optimizing the number of hidden layers, the number of neurons in these layers, the choice of activation function, learning algorithm, as well as the values of learning parameters of the selected algorithm. The developed ANN/GA method was implemented in the form of DEVONNA software agent that was validated through a case study Grancarevo, on the Tebisnjica river, in the Republic of Srpska, and the results were compared to the results obtained using the equivalent MLR/GA hybrid. Completed tests showed that the models generated by ANN/GA hybrid could give predictions of structural behavior of the dam with a higher accuracy than the MLR model. However, unlike the models in the form of MLR, which are resistant to temperature phase offsets present at different geographical locations, the models in the form of ANN exhibit unstable behavior under such circumstances. In addition, the generation of an ANN model has shown much higher computational demands. The comparative analysis of the models generated by the MLR/GA and ANN/GA methods on the one hand, and the models in the form of stepwise regression, on the other hand, has shown that the presented methods in some aspects surpass the capabilities of existing methods for generating the regression models. With the application of the research space dimension reduction the proposed hybrid methods and the developed software agents represent a powerful tool for modeling real objects and systems.