MULTI FORMULATED REGRESSION SLIGHTLY OUTPERFORM BACK PROPAGATION ARTIFICIAL NEURAL NETWORK ON RECOGNISING GAUSSIAN RANDOMIZED TWO DIMENSIONAL DATA AS BLOOD GLUCOSE LEVEL NON INVASIVE MEASUREMENT MODEL

Diabetes menyebabkan 1,5 juta kematian pada tahun 2012. Glukosa darah yang lebih tinggi dari pada optimal, sebagai pendahulu diabetes, menyebabkan 2,2 juta kematian tambahan. Pemantauan tingkat glukosa darah penting untuk mengurangi masalah. Ini mengilhami Kelompok Kerja Biomarker Non Invasive menjalankan penelitian untuk mengukur tingkat glukosa darah dengan perilaku non invasif, yang bekerja dengan menggunakan metode berbasis spektrofotometri, dan mesin inferensi yang tepat diperlukan untuk mengenali respons spektral alat. Tujuan penelitian ini adalah untuk


INTRODUCTION
Diabetes caused 1.5 million deaths in 2012. Higher-than-optimum blood glucose, as the precursor of diabetes, caused an additional 2.2 million deaths [1]. Blood glucose level monitoring is important to mitigate the problem. This inspiring Non Invasive Biomarking Working Group run the research to measure blood glucose level in non invasive manners, which works using spectrophotometry based methods. And several known spectroscopic methods currently in the research are as described in [2], [3], such as Amplitude Modulated Ultrasound with Infrared Technique [4], [5], Dielectric Spectroscopy [6], Electrical Impedance Spectroscopy [7], [8], LASER Reflected Spectral Patterns [9], Near Infra Red [10], and Occlusion Spectroscopy [11]. Artificial Neural Network long known as a good inference engine to mine raw data, in which we have a set of weighted calculation or activation function nodes that update it weight based on training data [12]. Known training methods include Incremental [13], Batch [14], Reverse / Back Propagation [15], and Quick Propagation [16]. Known activation function includes Linear, Sigmoid, Sigmoid Symmetric, Sigmoid Stepwise, and Sigmoid Symmetric Stepwise [17]. FANN [18] is one of the prominent engine for artificial neural network for C language.
In 2016, James Phillips proposes methods to subsequently curve fit and surface fit set of data using A large collection of equations for Python 3 curve fitting and surface fitting that can output source code in several computing languages, and runs a genetic algorithm for initial parameter estimation, called PYEQ3 [19].
He did not particularly give a name for His methods, so in this manuscript, the authors called it multi formulated regression (MFT).
Whichever method to choose, it results are spectroscopic data that's needed to be inferred to yield measurement data, and the proper inference engine is needed to recognize the device spectral responses. Currently, Performance comparison data between them is not available, as per our

METHODS
This is a comparative study, done in Material Physics Laboratory, Physics Department, Bogor Agricultural University. We create 500 rows of datums consist of 2 input and 1 output, with certified Gaussian random number generator [20] Comparative study done in January 2017. Rooted means squared error (RMSE) used for comparing fitting performance. We use data with 2 input and four output as training and testing data for both methods. Statistical Tools Employed include ALGLIB [21], [22], Zunzun [23], R [24], RKward [25], and epiR [26].

RESULT AND DISCUSSION
Computer Terminal Used are one unit of Hewlett Packard Laptop, 4 * AMD A8-6410 APU with AMD Radeon R5 Graphics [27], with 6.8 GiB of RAM, with Bodhi Linux 16.04 [28] installed.
FannTool [29] sets at3 Layers in 2-6-1 format, yields Incremental Training, Elliot Hidden Activation Function and Linear Piece Symmetric Output Activation Function as best settings ( Figure  1). Using 500 000 Epoch, The lowest RMSE is 15.295 mg / dl ( Figure 2).  Note that, we test the data which the correlation already known, so the result becomes a matter of accuracy. Due to time constraint, we did not test further heavier form of ANN, such as 4 or 5 layered one We would like to thanks Non Invasive Bio-marking Working Group, Bogor Agricultural University for Their indispensable support for this research.