کارایی الگوریتم کرم شب تاب و الگوریتم ژنتیک در تخمین ظرفیت تبادل کاتیونی خاک با شبکه های عصبی مصنوعی

در دهه های اخیر اندازه گیری غیرمستقیم ظرفیت تبادل کاتیونی (cec) با استفاده از توابع انتقالی مختلف موفقیت آمیز بوده است. شبکه عصبی مصنوعی (ann) نسبت به روش های رگرسیون آماری دارای نتایج مناسب تری بوده ولی با داده های اندک کارایی بالایی نداشته و از سوی دیگر فاقد یک روش استقرایی جامع در انتخاب الگوریتم یادگیری شبکه و توقف در حداقل محلی است. در این راستا استفاده از الگوریتم های بهینه سازی ضروری به نظر می رسد. هدف از این تحقیق، ارزیابی کارایی الگوریتم های کرم شب تاب (fa) و ژنتیک (ga) در تخمین cec با استفاده از ann است. برای نیل به اهداف فوق 220 نمونه از منطقه گلفرج به صورت تصادفی برداشته شد. سپس مدل سازی با سه مدل ann، شبکه عصبی مصنوعیژنتیک (ann-ga) و شبکه عصبی مصنوعیکرم شب تاب (ann-fa) انجام شد. در این تحقیق شبکه های عصبی با ساختار پرسپترون چندلایه، با الگوریتم پس انتشار خطا، تابع آموزشی بایزین و تابع محرک سیگموئید آکسون با 5 نرون مناسب ترین ساختار بوده است. نتایج نشان داد که مدل ann-fa دارای بیشترین کارایی بوده، به‌طوری‌که ضریب تبیین و میانگین انحراف مربعات خطا و ضریب نش ساتکلیف به ترتیب در مرحله آموزش 94/0، 31/1 و 53/0 و در مرحله آزمون 97/0، 06/1 و 59/0 بوده و مدل ann-ga در مقام دوم از نظر کارایی بوده است. میانگین هندسی نسبت خطا 84/0 برای مدل ann-fa بوده که نشان دهنده بیش برآوردی نسبی آن است. نهایتاً مدل پیشنهادی برای تخمین ویژگی خروجی مناسب بوده و کاربرد الگوریتم بهینه سازی کرم شب تاب و ژنتیک، حاکی از کاربردی بودن این الگوریتم ها در فرآیندهایی با طبیعت پیچیده و غیرخطی است.

Firefly Algorithm and Genetic Algorithm Performance in Cation Exchange Capacity Prediction by Artificial Neural Networks

Introduction: Cation Exchange Capacity (CEC) refers to the amount of negative charges available on the soil colloids surface. Clay and organic colloids carry a negative charge on their surfaces. Cations are attracted to the colloids by electrostatic bonds. Therefore, the charge of the soil is zero. For fertility map planning and commonly indicator of soil condition, CEC is an essential property. CEC is commonly measured on the fine earth fraction (soil particles less than 2 mm in size). CEC could be obtained directly but its measurement is difficult and expensive in the arid and semiarid regions with high amounts of gypsum and lime. Pedotransfer functions (PTFs) are appropriate tools to estimate CEC from more readily measured properties such as texture, organic carbon, gravel, pH and etc. Regression PTFs, artificial neural networks (ANN), and hybrids technique (HA) could be used to developing pedotransfer functions. The prior research revealed that could provide superior predictive performance when developed ANN model. Furthermore, ANN technique has no comprehensive method to select network learning algorithm and stopping algorithm in the minimum local. Therefore, application of optimization algorithms such as Genetic (GA) and firefly (FA) is necessary. The Purpose of the present study was to evaluate the performance of FA and GA to predict the soil cation exchange capacity by ANN technique based on easilymeasured soil properties.Materials and Methods: 220 soil samples were collected from 39 soil profiles located in Golfaraj (Jolfa) area of East Azarbaijan province. The study site lies from 45° 30ʹ to 45° 53ʹ east longitudes and from 38° 42ʹ to 38° 46ʹ north latitudes. Then, soil samples were airdried and passed through a 10 mesh sieve for removing gravels and root residues. Soil textural class, organic matter content and CEC were, respectively, determined by hydrometer, Walkley and Black, and bower methods. The artificial neural network (ANN), artificial neural networkGenetic algorithm (ANNGA) and artificial neural networkFirefly algorithms (ANNFA) models were applied to predict the soil cation exchange capacity on the basis of the easilymeasured soil properties. In ANNGA and ANNFA models, soil CEC was estimated via an artificial neural network and were then optimized using a genetic algorithm and firefly algorithm. The Genetic algorithms are commonly used to generate highquality solutions to be optimized by relying on crossover, mutation and selection operators. The firefly algorithm is modeled by the light attenuation over fireflies’ mutual gravitation, instead of the phenomenon of the fireflies light. The schema of flashes is frequently unique for specific types. The techniques’ results were then compared by four parameter, i.e., correlation coefficient (R2), root mean square errors (RMSE), Nash–Sutcliffe (NES) and Geometric mean error ratio (GMRE). Results and Discussion: The correlation coefficients of soil characteristic factors with CEC were analyzed through correlation matrix analysis. According to this analysis, the factors which had insignificant influence on the CEC were excluded. The clay, silt, sand and organic matter content were selected as input data. The parameter of the best deployment for MLP network could be used to predict CEC in the studied site. This model comprised 4 neurons (sand, silt, clay percentage and OM) in input layer. The optimum number of neurons in hidden layer was estimated to be 5. Additionally, the most efficient activity function in hidden layer was Axon sigmoid. Results showed that three CEC models performed reasonably well. ANNFA model had the highest R2 (0.94), lowest RMSE (1.31 Cmol+ Kg1) and highest Nash–Sutcliffe coefficient (0.53) in training stage and high R2 (0.97), lowest RMSE (1.06 Cmol+ Kg1) and highest Nash–Sutcliffe coefficient (0.59) in test stage. ANNGA model had also higher R2 (0.91), lower RMSE (1.77 Cmol+ Kg1) and higher Nash–Sutcliffe coefficient (0.45) in training stage and higher R2 (0.93), lower RMSE (1.50 Cmol+ Kg1) and higher Nash–Sutcliffe coefficient (0.48) in test stage indicating good performance of the model as compared with ANN models. The results showed that both ANN and Hybrid algorithm methods performed poorly in extrapolating the minimum and maximum amount of CEC soil properties data. In addition, the comparison of ANNFA, ANNGA results with ANN models revealed that ANNFA was more efficient than the others.Conclusions: The results of present study illustrated that ANN model can predict CEC with acceptable limits. Therefore, FA and GA algorithms provide superior predictive performance when is combined with ANN model. Firefly algorithm as a new method is utilized to optimize the amount of the weights by minimizing the network error. Final results revealed that this suggested technique improves the modeling performance.