Implementation of Neural Networks in Predicting the Understanding Level of Students Subject

This paper implements artificial neuralnetworkin predictingthe understanding level ofstudent’scourse. By implementing artificial neural network based on backpropagation algorithm, an institution can give a fair decision in prediction level of students' understanding of particular course / subject. This method was chosen because it is able to determine the level of students' understanding of the subject based on input from questionnaires given. The study was conducted into two ways, namely training and testing. Data will be divided into two parts, the first data for the training process and the second reading data of the testing process. The training process aims to identify or search for goals that are expected to use a lot of patterns. Thus, it will be able to produce the best pattern to train the data. After reaching the goal of training which is based on the best pattern, then it will be tested with new data to seeat the accuracy of the target data using Matlab 6.1 software. The results show that it can accelerate the process of prediction of students' understanding. By using architectural models 6-50-1 as the best model, some architectural models are tested and the result of prediction is reach to 87.75%. In other word, this model is good enough to make predictions on the level of students' understanding of the subject


Introduction
Private collegelike AMIK Tunas BangsaPematangsiantar, North Sumatera, Indonesia has not introduced a system of determining the level of students' understanding of the subject based on professional teaching activities. The process of learning is one of the activities in a college / institution in academic life [1]. The role is inseparable from professional teaching activities. In the learning process there needs to be a two-way relationship between students and professional teachers [2]. This meant that there was good cooperation during the learning process takes place. Evaluations are conducted by the university / institution of the learning process is necessary to end the semester. It is intended that no assessment of students and faculty professional. For these students aimed at assessing the level of understanding and absorption of the subjects that are taught and for faculty professional aims to assess the extent to which professional teaching activities can be channeled towards the subjects that diampunyapengetahuanya for 1 (one) semester. Therefore, the university or higher institution can provide equitable policies.
Prediction of students' understanding of the subject by professional teaching staff can help the university or higher institution in making policies on the future of professional teaching activities [3,4]. To determine the level of students' understanding of the subject is done by assessing (scoring) through questionnaires. Questionnaires were administered to represent the entire process of assessing the level of understanding of students on the course. Having obtained the target or goal of the assessment is desired then made the determination (determination) for the level of students' understanding of the subject by professional teachers. Artificial Neural Networks (ANN) is one of the information processing systems that are designed to mimic the way the human brain works in resolving a problem with the learning process through changes in weight [7][8][9][10][11]. There are many techniques that can be used for the implementation of Artificial Neural Networks one of them is Backpropagation [12][13][14][15]. Artificial Neural Networks by using the backpropagation algorithm has been widely used to solve some problems, one prediction problem.
This paper implements artificial neuralnetworkin predictingthe understanding level ofstudent'ssubject. This method was chosen because it is able to determine the level of students' understanding of the subject based on input from questionnaires given. The study was conducted into two ways, namely training and testing. Data will be divided into two parts, the first data for the training process and the second reading of the testing process. The training process aims to identify or search for goals that are expected to use a lot of patterns, then it will be able to produce the best pattern to train the data. After reaching the goal of training which is based on the best pattern, then it will be tested with new data to see at the accuracy of the target using Matlab 6.1 software.
The rest of this paper is organized as follow. Section 2 presents rudimentary of artificial intelegence, artificial neural networks and backpropagation algorithm. Section 3 presents experimental design. Section 4 presents results and following by discussion. Finally, the conclusion of this work is presented in Section 5.

Artificial Intelegence
AI is a field of study based on the premise that intelligent thought can be regarded as a form of computation -one that can be formalized and ultimately mechanized. To achieve this, however, two major issues need to be addressed. The first issue is knowledge representation, and the second is knowledge manipulation [1]. The main aim of Artificial Intelligence (AI) is to study how to build artificial systems that perform tasks normally performed by human beings. This concept was introduced in 1956 in the Darthmounth conference. From that moment on a lot of effort has been made and many goals have been achieved but unfortunately many failures as well. Today, the AI is a very important discipline and it includes a number of well-recognized and mature areas including Expert Systems [16][17][18], Fuzzy Logic [19][20][21][22], Genetic Algorithms [23][24][25], Language Processing, Logic Programming, Planning and Scheduling, Neural Networks and Robotics [26]. The general problem of simulating intelligence has been simplified to specific sub-problems which have certain characteristics or capabilities that an intelligent system should exhibit. Classical AI approaches focus on individual human behavior, knowledge representation andinference methods. Distributed Artificial Intelligence (DAI), on the other hand, focuses on socialbehavior, i.e., cooperation, interaction and knowledgesharing among different units (agents). Theprocess of finding a solution in distributed resolution problems relies on sharing knowledge aboutthe problem andcooperation among agents. It was from these concepts that the idea of intelligentmulti-agent technology emerged. An agent is an autonomous cognitive entity which understandsits environment, i.e., it can work by itself and it has an internal decision-making system that actsglobally around other agents. In multi-agent systems, a group of mobile autonomous agentscooperate in a coordinated and intelligent manner in order to solve a specific problem or classesof problems [27].

Artificial Neural Networks (NN)
Artificial Neural Network (ANN) is a computational model, which is based on BiologicalNeural Network. Artificial Neural Network is often called as Neural Network (NN) (See Figure 1). From Figure 1, to buildartificial neural network, artificial neurons, also called as nodes, are interconnected [7,8]. Thearchitecture of NN is very important for performing a particular computation. Some neurons arearranged to take inputs from outside environment. These neurons are not connected with eachother, so the arrangement of these neurons is in a layer, called as Input layer. All the neurons ofinput layer are producing some output, which is the input to next layer. The architecture of NNcan be of single layer or multilayer. In a single layer Neural Network, only one input layer andone output layer is there, while in multilayer neural network, there can be one or more hiddenlayer. An artificial neuron is an abstraction of biological neurons and the basic unit in an ANN [9,10]. The Artificial Neuron receives one or more inputs and sums them to produce an output. Usually the sums of each node are weighted, and the sum is passed through a function known as an activation or transfer function. The objective here is to develop a data classification algorithm that will be used as a general-purpose classifier. To classify anydatabase first, it is required to train the model. The proposed training algorithm used here is aHybrid BP-GA [11,12]. After successful training, user can give unlabeled data to be classified.

Figure 1. ANN Model
The synapses or connecting links: that provide weights, wj, to the input values, xj for j = 1 ...m. An additional function that sums the weighted input values to compute the input to the activationfunction as follow: where, w0 is called the bias, is a numerical value associated with the neuron. It is convenient to thinkof the bias as the weight for an input x0 whose value is always equal to one, so that; An activation function g: that maps v to g(v) the output value of the neuron. This function is amonotone function. The logistic (also called the sigmoid) function g(v) = (ev/(1+ev)) as theactivation function works best. The practical value of the logistic function arises from the fact that it is almost linear in the range where g is between 0.1 and 0.9 but has a squashing effectonvery small or very large values [28].

Architecture of Backpropogation
The back-propagation learning algorithm (BPLA)has become famous learning algorithms among ANNs. In the learningprocess, to reduce theinaccuracy of ANNs, BPLAs use the gradient-decent search methodto adjust the connection weights. The structure of a back-propagation ANN is shown in Figure 2. The output of each neuron is theaggregation of the numbers of neurons of theprevious level multiplied by its corresponding weights. The input values are convertedinto output signals with the calculations of activation functions. BackpropagationANNs have been widely and successfully applied in diverse applications, such as patternrecognition, location selection and performance evaluations [29].

Figure 2. Back-propagation ANN
There are several algorithms that can be used to create an artificial neural network, but the Back propagation was chosen because it is probably the easiest to implement, while preserving efficiency of the network. Backward Propagation Artificial Neural Network (ANN) use more than one input layers (usually 3). Each of these layers must be either of the following: a. Input Layer -This layer holds the input for thenetwork b. Output Layer -This layer holds the output data, usually an identifier for the input.
c. Hidden Layer -This layer comes between the inputlayer and the output layer. They serve as apropagation point for sending data from theprevious layer to the next layer [30].

Backpropagation Neural Network
Phases in Backpropagation Techniquealgorithm can bedivided into two phases: propagation and weight update.
Phase 1: Propagation Each propagation involves the following steps: 1. Forward propagation of a training pattern's input isgiven through the neural network in order to generate thepropagation's output activations.

Back
propagation of the output activations propagationthrough the neural network using the training pattern's targetin order to generate the deltas of all output and hiddenneurons. Phase 2: Weight Update For each weight-synapse: 1. Multiply its input activation and output delta to get thegradient of the weight. 2. Bring the weight in the direction of the gradient byadding a ratio of it from the weight.
This ratio impacts on the speed and quality of learning; itis called the learning rate. The sign of the gradient of aweight designates where the error is increasing; this is whytheweight must be updated in the opposite direction. The phases 1 and 2 arerepeated until the performance of thenetwork is satisfactory [31].

Evaluating the Performance of the Models
The main measures used for evaluating the performance ofmachine learning techniques for predicting the software effortare as follows [32]:

a. Sum Squared Error (SSE)
The sum squared error isdefined as.
where Pi = Estimated value for data point i; Ai =Actual value for the data point i; n = Total number of data points.

b. Mean Squared Error (MSE)
The mean squared erroris defined as.
where Pi = Estimated value for data point i; Ai =Actual value for the data point i; n = Total number of data points.

c. Root Mean Squared Error (RMSE)
The root meansquared error is defined as.
where Pi = Estimated value for data point i; Ai =Actual value for the data point i; n = Total number of data points.

d. Mean Absolute Error (MAE)
The mean absolute errormeasures of how far the estimatesare from actualvalues. It could be applied to any two pairs of numbers,where one set is "actual" and the other is an estimateprediction.
where Pi = Estimated value for data point i; Ai =Actual value for the data point i; n = Total number of data points.

Data Collection
In this study, the pattern recognition system and prediction of students' understanding of the subject is presented. Datawere obtained from questionnaires distributed to students grouped by subjects that taught by faculty as a lecturer. As for the format of a questionnaire given to students in the form of questions by 30 questions in which each question represents the study variables (See Figure 3). The research variables are learning, skills, assessment and workload, guidance and counseling, learning resources and standards and targets. Students need to give a score of 1=Strongly Disagree, 2=Disagree, 3=moderate, 4=Agree, 5=Strongly Agree, such as the following format:  The charging process is done by taking a sample of some of the classes are categorized by subjects that taught by professional teacher. The list of criteria to determine the prediction of students' understanding of the subject is given in Table 1 as follows:

Data Processing
Data processing is done with the help of matlab 6.1 software application. The raw data from a sample of 80 students will be divided into two parts, namely the test data and data testing. The raw data will be transformed into a table predetermined criteriain Table 2. As for the sample of data that has been processed and ditranformasi are as follows.

Table 2. Samples of Data that has been Transformed
Once the raw data is transformed into a table predetermined criteria, the next step is to determine the best architecture with patterns by performing a series of tests. The second stage is to predict the best architectural patterns obtained in the first stage. Testing process is carried out by entering the data and comparing the minimum error value is obtained from the best architectural pattern carried on the first stage.

Manual Design of Architectural Patterns
The desired outcome at this stage is the detection of a pattern determining best value for the level of students' understanding of the subject. The results are as follows: a. To determine the level of students' understanding of the course is based on professional teaching activities. Output from pemahan level there are 2 possibilities, namely "Understand" with weight 1 and "Not familiar" with weights0. b. Categorization "understand" or "do notunderstand" Category for "understand" is determined by the minimum error rate of the target "understand" that is 1. Those categorizations can be seen in Table 3 as follow: Categorization "not understand" is determined by the minimum error level of 0. The targets "do not understand" that "do not understand" the categorization can be seen in Table 4 as follow: International Journal of Software Engineering and Its Applications Vol. 10, No. 10 (2016)

The Best Pattern Determination
Training and testing is done several times with different parameters to get the best results with the software application Matlab 6.1 Neural Network method in determining the best pattern for the level of students' understanding of subjects by professional teaching staff has three part process, namely: a. The process of data input and the target includes inputting learning, skills assessment and workload, guidance and counseling, learning resources and standards and targets. As the target is the level of students' understanding of the course b. The process of determining the results of processed data that includes data conversion process into a pre-determined weight, calculate the weight value into the back propagation stage.
c. The results of the process of determining the data that is processed with Matlab 6.1 software application will be used to predict the level of student understanding terhadapat subjects by comparing the value of the minimum error.

ActivationfunctiontoHiddenLayer
: Tansig  To see the results of the comparison data MSE all architectural models are tested using matlab 6.1 applicationis complete, it can be seen from Table 6 as follow:

Conclusion
In this paper, implementation of artificial neural network in predicting the understanding level of student's subject has been presented. Based on the results and analysis of the previous chapter, the author can draw conclusions as follows: a. Adding lots of hidden layer during the training and testing, not a maximum results. To 5 models designed architecture, 6-50-75-1 is a model that has the largest MSE is 0.028540763 b. Having carried out experiments in the training process and system testing is done using Matlab 6.1 software application. Neural Network Model used was 6-50-1, 6-75-1 models, the model 6-100-1, 6-50-75-1 models and models 6-75-100-1, can be obtained good results with a view MSE the smallest and fastest epochsis 6-50-1.
c. With 6-50-1 architectural models, can make predictions on the students' understanding of the subject by showing the performance above 92%.