A Genetic Optimized Parallel MLP to Improve Classification Accuracy for Web Learning System L. Jayasimman 1, E.George Dharma Prakash Raj 2 1 Department of MCA, JJ College of Engineering and Technology, Trichy, Tamilnadu, India 2 School of Computer Science and Engineering, Bharathidasan University, Trichy, Tamilnadu, India Abstract: Web-based learning system helps to improve the style of learning environment that adapts the learning material to meet learners needs and also can adapt the learning content according to the individuality of learners. Web learning system provides personalized learning environment that is based on user s cognition. Web Based Education-WBE encompasses all aspects and processes of education that use a World Wide Web as a communication medium and supporting technology. With the increasing popularity of web based learning, it is required to recognize the cognition of individual learners. Cognitive theory is widely used to predict the effectiveness of the web learning and multimedia learning. The cognition induced by instructional and multimedia modes are measured by indirect or subjective methods. Questionnaires are a common form of measuring cognition indirectly. In this paper, a questionnaire is prepared to identify the cognition of the student and his website preferences in the web learning environment. Then all the user preferences are classified using proposed neural network classifiers with an optimization technique. The cognitive attributes are used as the training input for the proposed genetically optimized neural network. Keywords: Web Learning System, Classification Accuracy, Genetic Optimization, Parallel Multi Layer Perceptron, Neural Network. 1. INTRODUCTION The real value of Web-based learning lies not in accessing knowledge at any time, any place, and for anyone, but helping the right students to acquire the right skills and knowledge at the right time in order to function as active, self-reflected and collaborative participants in the information based society. Only the Web-based learning becomes a crucial resource for learners and educational institutions[1]. This however can be achieved only by adhering to the learning paradigm and associated. pedagogical principles, and to the factors that constantly affect the development of Web-based learning. Unfortunately, much of the development of Web-based learning is carried out without a true understanding of the issues that are proper to Web-based learning, partly because marketing advertising and technologies still drive the construction process [2]. Clearly, there is a need for a disciplined, systematic approach to the development process for incorporating and translating the specific requirements of Web-based learning into a system that must constantly evolve in order to ensure the relevance, correctness, and completeness of the content available on the Web[3]. Cognitive learning theory [4], states that learning involves a change in a person s cognitive structure, this change occurring when new information/experiences combine with existing knowledge stored in the long term memory (LTM). Learning is meaningful when it connected to what a person already knows. When new information is connected to existing knowledge, it is known as meaningful learning. Effective teachers create learning experiences leading to such learning. Classification is a data mining technique that assigns items in a collection to target categories or classes. The goal of classification is to accurately predict the target class for each case in the data. Neural networks have emerged as an important tool for classification. The recent vast research activities in neural classification have established that neural networks are a promising alternative to various conventional classification methods[5]. The advantage of neural networks lies in the following theoretical aspects. First, neural networks are data driven self-adaptive methods in that they can adjust themselves to the data without any explicit specification of the functional or distributional form of the model. Second, they are universal functional approximators in that neural networks can approximate any function with arbitrary accuracy. The relevance in optimization technique will lead to the genetic approach which has a little chance to get stuck in local minima. Genetic approaches are finding prevalent applications in solving problems requiring efficient and effective search in the synthesis of neural network architectures, scheduling, numerical optimization etc., and results in solutions that are globally optimal or nearly so. Topology optimization, genetic training algorithms and control parameter optimization are the ways in which genetic algorithms are applied in Artificial Neural Network (ANN). In Genetic algorithm, the learning of an Volume 2, Issue 4 July August 2013 Page 315
ANN is formulated as a weight optimization problem such as learning rate, momentum rate and the tolerance level is optimized [6]. A great deal of research is going on in the neural networks worldwide and its applications are expanding in various domains and disciplines such as character recognition, image compression, stock market prediction, medicine and many more. Backpropagation has been used in recognition and learning in neural networks. Neural network classifiers have been shown to provide supervised classification results that significantly improve on traditional classification algorithms such as the Bayesian Maximum Likelihood (ML) classifier. The feed forward multilayer perceptron is the predominant neural network architectur. When a new method was used for interpreting the weights of a trained network, it was proven that the neural networks are able to adjust their weights in accordance with the importance of the role each input data source plays during the classification.the use of genetic algorithm in the design of ANN improves the classification Accuracy. Yung et al. [7] predict the stock price using a hybrid Genetic Approach (GA) combined with recurrent NNs describes a number of input variables that help the network to forecast the next day price. For the input variables, technical indicators or signals are used that were developed in deterministic trading techniques. The backpropagation algorithm is prone to get stuck in local minima and highly depends on the initial weights; to avoid this GA is used for optimizing NN s weights. This paper is organized into the following sections. Section 2 briefly describes the related works, section 3 discusses the methodology and genetic algorithm, section 4 describes the results obtained and discusses the same. Finally the section 5 concludes the paper. 2. RELATED WORKS Jan L. Plass [9] proposed a hybrid model that combines cognitive and software engineering approaches regarding the criteria for the design and evaluation of the user interface of foreign language multimedia software. The proposed approach involves a three step design which includes selection of instructional activity that supports cognitive processes of competence, selection of feature attributes and selection of design features. It is still pragmatic to be practical. Based on this proposal, contextualized model of interface design, domain specific evaluation criteria are developed to describe how well the user interface is able to support the cognitive processes involved in the development of linguistic and pragmatic skills and competencies in SLA. Baylari et al [8] proposed a personalized multi agent e- learning system based on item response theory (IRT) and artificial neural network (ANN) which presents adaptive tests (based on IRT) and personalized recommendations (based on ANN). These agents add Adaptivity and interactivity to the learning environment and act as a human instructor which guides the learners in a friendly and personalized teaching environment. The framework is constructed adaptive tests that will be used as a posttest in the system. Thus a multi-agent system is proposed which has the capability of estimating the learners ability based on explicit responses on these tests and presents him/her a personalized and adaptive test based on that ability. Also the system can discover learner s learning problems via learner responses on review tests by using an artificial neural network (ANN) approach and then recommends appropriate learning materials to the student. Experimental results showed that the proposed system can provide personalized and appropriate course material recommendations with the precision of 83.3%, adaptive tests based on the learner s ability, and therefore, can accelerate learning efficiency and effectiveness. Urszula et al. [10] proposed GenPar method of a Neural Network (NN) description in classification problems by means of genetic algorithms. For that reason NN produces the training examples for developed method. In the GenPar two genetic operations are implemented. That is crossover and mutation. Because the individual is a list of genes with variable length, the operators differ from their standard form. Individuals are selected by applying the roulette wheel method because individuals have different genotype length. The work assures that both parents have an equal fine contribution in the offspring chromosome. During the crossover, the offspring chromosome is created by a random choice of a gene from one of the parents till to the length of shorter parent. There are two forms of mutation. The first one, on the higher level, adds or deletes a gene representing a single rule. On the second level it is realized as a random change of binary values like: flag or a bit in the binary sequence for enumerative parameters. Limit values of real attributes, which are implemented as 32 bits numbers, are mutated in the same way. Urszula, [15] gives an appropriate definition of neural network architecture prior to data analysis is crucial for successful data mining. This can be challenging when the underlying model of the data is unknown. The goal of this study was to determine whether the optimizing neural network architecture using genetic programming as a machine learning strategy would improve the ability of neural networks to model and detect nonlinear interactions among genes in studies of common human diseases. Furthermore, the genetic programming optimized neural network is better than the traditional back propagation neural network approach in terms of predictive ability and power to detect gene-gene interactions when non-functional polymorphisms are present. 2.1 Genetic Optimization Techniques Genetic Optimization is the evolutionary approach which aims to improve the solution to a problem by keeping the best combination of input variables. ANN is the best tool for classification and regression. The parameters of the Volume 2, Issue 4 July August 2013 Page 316
ANN are optimized with a genetic approach to enhance the classification accuracy. When the ANN is combined with genetic approach, is known as hybrid model which is better performed than the conventional ANN. The Genetic Approach often works with a form of binary coding. If the problems are coded as chromosomes, the populations are initialized. Once the population size is chosen, the initial population is randomly generated. After the initialization step, each chromosome is evaluated by a fitness function. According to the value of the fitness function, the chromosome associated with fittest individuals will be reproduced more often than those associated with unfit individuals. 2.2 Cross-over Cross-over means mating between individuals [11]. A pair of chromosomes is picked up at random and the single-point crossover operator is applied according to a fixed cross-over probability. For this operation, a random number in the range of 0 to the length of the string is generated. This is called the crossover point. Two portions of the strings lying to the right of the crossover point are interchanged to yield two new strings. Mating is the creation of one or more offspring from the parents selected in the pairing process. The most common form of mating involves two parents that use a crossover to produce two offspring. The k-point and flat crossover are among the commonly used crossover techniques. Inkpoint technique, k crossover points within the range of an individual are selected randomly and the individual parts between these points are swapped by the parts of another individual. In this paper, we use two-point crossover technique, where the points are selected, randomly. Using the two-point crossover, each of the parents is divided into three parts. The second part of the individuals is swapped and two offspring is formed. 2.3 Mutation Mutation introduces random modifications [11]. Mutation consists of considering in turn each bit of a given chromosome and changing its value with a predefined low probability called the mutation. After mating process, a part of the offspring is mutated. Random mutations alter a certain number of bits in a mutant offspring. The single point mutation which changes a bit one to zero and vice versa. Mutation points are randomly selected from the Di N bit number of bits in the offspring array. Mutation rate influences convergence of the algorithm. On one hand, large mutation rate increases the diversity of the algorithm which is considered as a good way for avoiding premature convergence. On the other hand, it tends to distract the algorithm from converging on a popular solution. Hence, an appropriate mutation is always needed. 2.4 Selection Selection equates to the survival of the fittest. Fitness is determined by an objective function or by a subjective with lower fitness die and individuals with higher fitness survive [Siv, 11]. 3. METHODOLOGY 3.1 Multi Layer perceptron ANN architecture is also known as a multilayer perceptron (MLP). Neural networks operational principle is simple. Each input layer neuron has a value, so that the input layer holds input vector. Each neuron connects to other neurons in the next neuron layer. Artificial Neural Network s architecture is a neuron layout grouped in layers. ANN s main parameters include: layer numbers, neuron number per layer, connectivity level and neurons interconnector types. A multilayer Perceptron (MLP) is an original Perceptron model variant proposed by Rosenblatt in 1950 [13]. It has one/more hidden layers between input and output layers, neurons are in layers, connections are always directed from lower to upper layers, same layer neurons are not interconnected. The neural network s first layer is the input layer containing n neurons; the last network layer is the output layer, containing m neurons. In the Perceptron model, a neuron with a linear weighted net function and threshold activation function are used. Input to a neuron is a feature vector in n- dimensional feature space. The net function is a weighted sum of inputs: 3.2 Input Layer A predictor variable vector of values is presented to the input layer. The input layer distributes values to each neuron in the hidden layer. In addition to the predictor variables, a constant input of 1.0, called bias is fed to each hidden layer; the bias is multiplied by a weight and added to sum going into the neuron [11]. 3.3 Hidden Layer Neurons between an input and output layers are the hidden layer neurons. Outputs from hidden layer are distributed to the output layer. First hidden layer neurons are directly connected to input layer (data layer) of neural network. 3.4 Output Layer Reaching an output layer neuron, value from every hidden layer neuron is multiplied by a weight (w kj), and resulting weighted values are added, producing a combined value V j. The weighted sum (V j) is fed to a transfer function σ, which outputs value Y k. The Y values are network outputs. The logistic function defined by: Volume 2, Issue 4 July August 2013 Page 317
4) Mutation: probability of inherited gene mutation in children was considered 0.1 (gene value is modified to random value). 5) Generation of new Population: The previous population s fittest chromosomes and 6 children generated form a new population for the next generation. 6) Steps 2 to 6 are performed till the number of iterations corresponds to predefined time or maximum generations. Figure 7 Mapping the weights of the neural network from Problem space (left hand side) into a chromosome. 3.5 Genetic Algorithms (GAs) GAs finds approximate solutions to difficult problems through application of evolutionary biology principles to computer science. GAs views learning as a competition between the populations of evolving candidate problem solutions. A fitness function evaluates every solution to decide if it can contribute to next generation solutions. By operations analogous to gene transfer in sexual reproduction, the algorithm generates a new candidate solution [14] problem. The 3 important aspects of using GA are: Definition of objective function. Definition/implementation of genetic representation and Definition/implementation of genetic operators. begin t 0 initialise P(t) evaluate P(t) while(not termination condition) do begin t 0 select P(t) from P(t 1) alter P(t) evaluate P(t) end 3.7 Steps for the genetic algorithm [12] 4. EXPERIMENTAL RESULTS The cognitive behavior of 82 students studying in undergraduate and postgraduate courses was captured using questionnaires. They were initially subjected to go through a known subject and an unknown subject in a popular online learning website [14]. The typical questions were in the areas of Learning ability Indication about meaningfulness of error Message Prefer to read the text rather than to listen to a lecture Visualization of content read as a mental picture Typical questions in the questionnaire are as follows: 1. I prefer content that is challenging so I can learn new things. 2. Compared with other websites this website is better in terms of content. 3. I am so nervous during the online test that I cannot remember facts I have learned 4. I often choose advanced concept links even if they require more work 5. I am sure I can do an excellent job on the problems and tasks assigned for this session 6. I think I will receive a good grade in this class 7. Even when I do poorly on a test I try to learn from my mistakes 8. I think that what I am learning in this class is useful for me to know 9. I think that what we are learning in this website is interesting 10. Understanding this subject is important to me Table 1 Neural network parameters 1) Generation of Population: Initial population generated with 10 chromosomes and population size maintained over generations. 2) Evaluation: Each chromosome evaluated to find fitness. Parent s selection: Here among top 5 chromosomes with better fitness, 4 are chosen randomly to an empty parent chromosome set. Among the 5 with worst fitness, 2 are chosen to the same set. 3) Passage of genes from parents to children: The principle is that identical characteristics between parents should pass on to children. Volume 2, Issue 4 July August 2013 Page 318
Table 2 Genetic algorithm parameters Initial population size 20 Maximum generations 20 Number of epochs 500 Momentum optimization Lower bound 0.5 Upper bound 1.0 Step size optimization Lower bound 0.1 Upper bound 0.5 Encoder mechanism Roulette Cross over Single point Cross over probability 0.75 Mutation Uniform Mutation probability 0.01 Table 3 Classification Accuracy Neural Network General Cognition Algorithms MLP 75.6% Parallel MLP 92% Proposed GO PMLP 92.68% The above result shows that GO PMLP gives better classification accuracy. 5. CONCLUSION This study considered a new method to identify and clasifiy the users preferences of web learing system based on his/her cognitive behavior. It was proposed to optimize the weight updating parameters of PMLP (GO PMLP) and improve the classification accuracy. MLP is combined with GA gives better classification accuracy. Questionnaires were used to indentify the cognitive behaviour of the students. The proposed Genetic Optimized PMLP achieves a classification accuracy of 92.68% for web learning data set. References [1] Joi L. Moore, Camille Dickson-Deane, Krista Galyen, e- Learning, online learning, and distance learning environments: Are they the same?, Elsevier Inernet and Higher Education, Volume 14, Issue 2, pp. 129-135, 2011. [2] Thavamalar Govindasamy, Successful implementation of e-learning Pedagogical considerations Internet and Higher Education, Volume 4, Issue 2, pp. 287 299, 2002. [3] Ivan Martinez, Pablo Moreno-Ger, Jose Luis Sierra, Baltasar Fernandez-Manjon, Educational Modeling Languages: A Conceptual Introduction and a High Level Classification, computers and education E- learning, From theory to practice, Springer, pp. 27-40, 2007. [4] Andy Johnson, Applying Cognitive Learning Theory in The Classroom, pp. 1-7, 2012. [5] Peter Zhang, Neural Networks for Classification: A Survey IEEE transactions on systems, man, and cybernetics part c: applications and reviews, Volume 30, Number 4, pp. 451-462, 2000. [6] Fausto Pedro, Garcia Marquez, Marta Ramos, Martin Nieto, Recurrent Neural Network and Genetic Algorithm Approaches for a Dual Route Optimization Problem: A Real Case Study, Proceedings of the Sixth International Conference on Management Science and Engineering Management, Lecture Notes in Electrical Engineering-Springer- Verlag, Volume 185, pp. 23-37, 2013. [7] Yung-Keun Kwon, Byung-Ro Moon, A Hybrid Neurogenetic Approach for Stock Forecasting, IEEE Transactions on Neural Networks, Volume 18, Issue 3, pp. 851-864, 2007. [8] Ahmad Baylari, Montazer, Gh. A., Design a personalized e-learning system based on item response theory and artificial neural network approach, Expert Systems with Applications, Volume 36, Issue 4, pp. 8013-8021, 2009. [9] Jan L Plass, Design and Evaluation of the User Interface of Foreign Language Multimedia software: A Cognitive Approach; Language Learning & Technology, Volume 2, Number 1, pp. 35-45, 1998. [10] Urszula, Rule extraction from neural network by genetic algorithm with Pareto optimization, Springer, pp 450 455, 2004. [11] S.N. Sivanandam, S.N. Deepa, Principles of Soft computing, Wiley India(P) Ltd., 2007. [12] Abdella, M., &Marwala, T., The use of genetic algorithms and neural networks to approximate missing data in database, IEEE 3rd International Conference on in Computational Informatics, pp. 207-212, 2005. [13] Reiterer, H., The development of design aid tools for a human factor based user interface design, International Conference on Man and Cybernetics Systems, pp. 361-366, 1993. [14] Mohamed Ettaouil, Mohamed lazaar, Youssef Ghanou, Architecture optimization model for the multilayer perceptron and clustering, Journal of Volume 2, Issue 4 July August 2013 Page 319
Theoretical and Applied Information Technology, Volume 47, Issue 1, pp. 64-72, 2013. [15] Urszula, GA-Based Rule Extraction from Neural Networks for Approximation, Proceedings of the International Multiconference on Computer Science and Information Technology, pp. 141 148. AUTHOR L. Jayasimman working as a Assistant Professor, with department of Computer Application, J.J.College of Engineering and Technology, Trichy, India. He received his M.Tech degree in Bharathidasan University, Trichy, India in 2008. Now he is pursuing PhD (Computer Science) in Bharathidasan University. Dr. E. George Dharma Prakash Raj working as a Assistant Professor with School of Computer Science and Engineering, Bharathidasan University, Trichy, India. He has more than twenty years of experience in Academics. He is a Chairman/Member in Board of Studies in Colleges and Universities in India. He is also a member in Research Organizations, International Programme committee member, Reviewer for international/national journals and conferences. Volume 2, Issue 4 July August 2013 Page 320