Automated Modelization of Dynamic Systems

Transcription

1 Automated Modelization o Dynamic Systems Ivan Perl, Aleksandr Penskoi ITMO University Saint-Petersburg, Russia ivan.perl, aleksandr.penskoi@corp.imo.ru Abstract Nowadays, dierent kinds o modelling settled down in most areas o human activities. Oten, it is not easy to design and create a model o an existing systems, because equations linking its components are not known. This article describes an approach allowing to identiy system components interaction based on run-time monitoring and analysis o system behavior. Described method is shown on a simple system that will be turned in to a system dynamics mode by automated equations deinition and results will compared to initial analytic model or the same system. I. INTRODUCTION We are living in a complex world and with new technological and scientiic achievements our world became even more complex, almost, on a daily basis. That is why modelling taking leading positions when it comes to question o systems analysis and design. There are many dierent classes o systems and, consequentially, number o ways to design, deine, build and compute related models. One o the most requent kinds o systems we are seeing around are dynamic systems, or systems, behaving or evolving over time. Approach to work with kind o systems is called System Dynamics. Appeared initially in mid 60s and designed or describing economics processes, later, in terms o growth o its methodological maturity, became applicable to work with almost any dynamic system. System dynamics models consist o two key elements: stocks and lows. Where stocks are representing accumulation o anything accounts receivable, capital, delays, inventories, and so on and lows are representing movement o anything consumption, deliveries, expenses, production, etc.[1] Today, computer modelling, like many other services, is ollowing a trend o moving to the cloud. There are many beneits o using cloud services, comparing to maintenance own deployments o required systems and services, but in a light o discussed topic, key o them are computation capacity and simpliied data access. In most cases, cloud services, especially designed or computation and modelling needs, are hosted on a specialized and well conigured equipment, which is ine-tuned or solving particular types o tasks. Because o various kinds o optimization s, such hosted solution will be more eicient even comparing with the same hardware hosted locally, but without speciic ine-tuning. That is why cloud services can do more and do this cheaper. Second aspect is data access. When dealing with cloud services there are no issues with bubbling data up to the client or with getting them back or sharing with colleagues and your research ellows. When it comes to systems modelling these two beneits starting to play even more valuable role together. Having eicient model execution engine, taken in parallel with monitoring data o the real systems allows to design and build unique services that can change systems analysis and management approaches, by shited these process to a kind o augmented reality, where real systems are extended by their models and their models, like digital shadows, became their integral part. One o such solutions, called continuous modelling, described in [10], allows to perorm monitoring o Internet o Things aware systems in parallel with their running models with constant comparison o expected state o the system with its actual state, what allows to quickly identiy anomalies and deviations in system behavior. Another joint processing o a model and real system data beneit can be achieved during creation o the model or an existing system that can be monitored in IoT-like way. When researcher is working on a new system or trying to deine a new process, he can deine both model structure (stocks, lows and their coniguration) and equations he is going to put into his model as lows deinitions. But not all works are starting rom scratch. System Dynamics is requently used to optimize and improve already existing systems which evolved over time without having their own models o any kind. In such cases, irst action item which rises in the beginning o the work is analysis o the system with identiication o its structure and understanding its behavior. For example, requent case is when we are having a running system and we need to understand how it behaves and what are the options we have to perorm its optimization. When we start dealing with such situation, in majority o cases we do not have equations which will describe interactions between system components. In this case, to create a model we will need to reconstruct required dependencies via system monitoring and analysis. II. MODELIZATION DEFINITION Today more and more system became a part o a paradigm called Internet o Things. In general, it means that components o such systems became connected to the Internet and can be managed and monitored using centralized services. New services, manuacturing, production and other systems became designed now to be IoT-enabled rom scratch and in the same time many systems which are working or years already are enhanced to be IoT-aware with wide range sensors and dierent eedback processing devices allowing inluence ISSN

2 Fig. 1 Basic structure o a demo model with two lows and three stocks systems remotely. An easy example is equipping production lines with monitoring sensors, and remote power management units allowing to start, stop and monitor production lines using a centralized service. Since new technologies are coming to scene, questions about new ways o system optimization are bubbling up. Now with all new monitoring possibilities, it is reasonable to try leveraging them to achieve better productivity and reduce various kinds o waste. Another aspect to take into account is growing computation power and computation accessibility. Today we can perorm much more computations than even ew years back and this leads to growing application o all kinds o computer modelling to support almost any human activity. Joining together these two eatures, gathering o IoT data and application o models or system understanding, monitoring and improvements, can bring lost o beneits or most o industries. In terms o this article, we will ocus on working with dynamic systems and under selected modelling approach we will assume System Dynamics toolkit. To illustrate uture discussion we will use a simple model that consists o three stocks and two lows. For example, such model may represents a supply chain pipeline where we have production warehouse, consumer one, that process items rom this warehouse and consumer two that perorm a inal processing over results o irst-line processing made by consumer one. Fig.1 schema depicts described model. In terms o this articles, under system modelization, we will understand a process o automated deinition o equations representing system dynamics model lows by analyzing stocks monitoring. The basic idea o this approach can be explained in the ollowing way: 1) Model structure deinition. During this step should be deined key stocks (nodes o the model) and lows. Here we are interested only in the structure o the model without its unctional aspects, and it will be enough to deine source and destination stocks or each low without speciying any equations explaining low behavior. 2) Data acquisition. At this stage, we will perorm monitoring o the target and will retrieve the state o each system node. As a result o this stage, we will get a list o aligned time series data sets describing a state o each model low over time. It is important to have this time series aligned in time with the precision that makes sense or the concrete use case. For example, i we are talking about supply chain model, several minutes drit between time series will not be critical, but i we deal with the model o a short running process like cooling down a teacup - several minutes will be the time o the whole experiment. 3) Monitoring data analysis. Ater all time series are received and aligned in time, they should pass through a chain o statistical processing methods like an approximation. The goal o this step is to ind an equation that will describe changes o stock (node) state over time or o low intensive. 4) Restoration o non-measured data. In practice, during the experiment, we can t get all data o a system, or all stacks and lows because technical or economic reasons can t measure some o them. The goal o this step is to ind an equation the will describe changes o stock state and low intensive which we can not measure directly. 5) Flows equations reconstruction. This step is a key element o the proposed schema. Taking into account the behavior o each model stock as unctions rom the time, we will try to reconstruct lows, which are linking stocks which behavior we just investigated and described analytically with equations we extracted rom system monitoring results. It is important to understand, that in most cases, it is not possible to achieve the goal o complete and ully automatic reconstruction o system dynamics data rom real system monitoring, especially or complicated systems with big numbers o lows going back and orth between stocks. Goal o the proposed approach is to design and develop an instrument that will help to reconstruct a model o the system specialist needs to work with, buy providing him advanced and enhanced view o the monitored system with the number o recommendations that will help him to solve this problem. Whole system modelization algorithm can be split on two phases: deining model structure, setting up monitoring, retrieving monitoring data and processing them to get analytic unction deinitions instead o pure time series. Second phase will be an attempt to reconstruct a model rom obtained and processed data. First part o this algorithm is sown on ig. 2 III. MODEL STRUCTURE DEFINITION This is a irst step, in building a model o existing system. When passing this step, it is important to understand that model complexity, will aect complexity o the rest o the process and higher will it be - more complicated computations will be required and more assumptions will be needed to be made on urther process stages. While working with model structure it is better to ollow incremental processes like ones described or sotware development in Agile methodologies. Process o model reconstruction can be treated as an incremental detalization and improvement

3 Deine basic model structure InFlow Stock OutFlow Deine system monitoring pattern Flows monitoring Stocks monitoring Hybrid Stock Value Fig. 4 Stock state monitoring options Get analytic deinitions or monitoring results Fig. 2: First phase o modelization algorithm Fig. 3. Simpliied structure o a demo model with two stocks and only one low o the model we are working on with step-by-step diving into the system under analysis. Getting back to our example with a production pipeline with two sections, in a very irst version o the model we may deine it at a glance like a two step model, like shown on ig. 3. In this case, whole processing will be treated as a one single operation without dividing it into processing stages. Such simple model will allow to understand overall system behavior. When on such level all required model components will be reconstructed, this model can made more complicated and intermediate processing stage can be revealed, like it was shown on initial ig. 1. Whereat can be revealed sub-steps o each processing step and so on, until model will not reach required detalization level to address issues it is designed or. Each deine d stock will require perorming monitoring and analysis o monitoring results. For each deined low we will need to pass all steps to reconstruct low equations rom the monitoring data retrieved rom related to it monitoring low. IV. DATA ACQUISITION Ater deining model structure, next important step is proper acquisition o model data. Since we are dealing with real systems, it is not always possible to properly measure states o all system components we are going to include in the model. There are two key measuring cases, how system component can be monitored: state monitoring and lows monitoring. In case o system dynamics model, model nodes, represented by stocks are not unctional and they can be treated just like a storage accumulating some entities. In the same time lows are processes which are changing values o the stocks. In other words, everything what perorms actions on stocks values is a low. For example, a truck which is moving a cargo rom one storage to another is a low connecting two warehouses (stocks). Another example is urniture actory, it consumes wood rom one place, produces urniture and puts it in another. In this case, actory is a low and on dierent sides o this low it c onnects two s tocks w ith c ompletely dierent entities. When we are perorming system monitoring, in most cases we can monitor only state o the stocks, but not lows directly. On ig. 4 shown two options or stock monitoring. I architecture o the system allows to measure stock input and output low, it m eans, t hat we aced a r elatively easy case, because during system monitoring we knows, direction (increasing or decreasing) and pace at which stock value is changed. This is already vary close to the solution we are trying to achieve - deinition o the lows in th e running system. In most cases, it is not possible to properly measure input and output lows and the only metric that can be monitored is stock actual value. In other words, lets say, that we have scales with sugar and there are two lows: one adding sugar to the scales and other is taking it rom there. These two processes are perorming simultaneously and the only thing we know at any given moment is amount o sugar on scales. In this case we will not be able to directly deine properties o each low. To this topic we will get back later in chapter?? V. MONITORING DATA ANALYSIS Next important step is data analysis. In the current context, under data analysis, we will assume processing o monitoring data we got rom the system under analysis. As a result o monitoring, we will receive a time series describing required system components which are represented by stocks in the model. From processing perspective it is not always easy to work with time series directly, also, taking into account that our goal is to ill gaps in the model deinition with illing in, it required to convert this data series into analytic representation. There are several dierent approaches allowing to ind a unction that will approximate experimental data. Approximation o experimental data is a topic that is well described in many sources like [4] or [5]. The goal o approximation is to deine an analytic unction that will go as close as possible to the set o experimental data. When approximation is perormed we need to do key things: irst, deine a unction that will be used or approximation, or example linear, polynomial, exponential, trigonometric and so on, and second - deine proper coniguration and coeicients o this unction. Oten, when a researcher or analytic receives experimental data, he can try guessing the best approximation unction by looking at experimental data charts. In other words, based on

4 his own unique experience, concrete case and a set o available data, he is making a decision about the shape o the unction, that will give him closest approximation results. When designing cloud service solution, it is not possible to address approximation aspect this way, because system under development should be tolerant to dierent sequences o input data received or analysis and these data sets should be processed with some deined precision and quality level. To address this problem in an automated way, there is a couple solution. The irst approach that can be used is a multivariate parallel approximation. For that purpose, on a platorm level, we will deine a set o most popular unctions that can be used or approximation purpose. For example, we may have linear, polynomial, exponential, logarithmic and trigonometric approximation bases in the library. When new data set received or analysis, the system will do a parallel approximation using each o bases available in the library. Quality o approximation can be measured, or example, by using standard deviation between experimental data and ound approximation unction. When all parallel approximations are completed, we can make a decision about best-ound approximation using selected approximation quality metric. The key beneit o such solution will be the ability to easily extend library o approximation bases with new entities and with the ability to update existing ones with better computation methods, because, the polynomial approximation can be perormed with the number o was, or example using a method o Least Squares or using Tailor series. The principal disadvantage o such an approach will be high consumption o computation resources. When designing a cloud solution, we have to assume, that there can be a signiicant number o researches, pushing data to the service or analysis and all o them are expecting responses in a reasonable time. The approximation is an expensive job rom resources consumption perspective and perorming it in parallel with a big library o approximation bases can give very good results, but will take much time. This method can be made cheaper, i we will allow speciying the number o patterns researcher expects in his data, this will lead to decreasing number o required computation, but will require knowledge about time series passed to or analysis. The second approach that might help to address the issue with deining closest approximation unction shape is an application o machine learning algorithms. Machine learning works well or dierent classiication jobs. Deinition o time series shapes is a typical classiication problem where we need to identiy the ranked set o most suitable shapes or the given time series. Deep learning has been used successully or time series prediction. In particular, recurrent neural networks (RNNs), especially those utilizing long short-term memory (LSTM) nodes, are useul or sequential tasks like this. Lipton, Kale, and others [6], have used LSTM RNNs or time series o medical data. They demonstrated that an RNN can ingest a sequence o clinical medical observations and then accurately classiy many diseases based upon these time series o raw sensor readings and lab tests. Having a library o expected time series shapes and trained neural networks will allow reducing the number o parallel approximations we need to attempt to get the result, closest to the sensor data processed. Let describe what we have ater this step rom the mathematical point o view: Functions (F, G, H...) o all measurement stocks rom the time (F : time value). Function o stock value change or a speciic time interval (F : time valuechange). Also, in non-strict means, we can call it a derivative o a stock unction. Functions o all measurement low inluence to source and destination stocks ( : time valuechange and 1 : time valuechange). It should also be noted that this unction is not a low equation, because the low is a unction o stock value, not rom the model time. VI. FLOWS EQUATIONS RECONSTRUCTION This is a key step o the described approach. Goal o this step is to attempt to recreate equations, describing each low o the model which structure was deined in the beginning o modelization process. Let ormalize a problem that should be solved on this step. We already have an inluence unction o most low in the model, but as noted above, it s not a low deinition, because low is a unction rom stocks values : (stockvalue1,stockvalue2...) valuechange. At this step we should replace all low inluence unction to it s low equivalent in regarding to the ollow term: (t) (F (t),g(t)...) There are number o options to use or lows unctions reconstruction and approach selection will depends on what kind o monitoring was possible to perorm on a system under analysis. 1) Input and output lows monitoring. Let s say that or a pair o stocks (or system components) we were able to perorm monitoring o output low or the source stock and input low o destination stock. It means that or this particular low we have its input and output values. Since low, in terms o model, is a pipeline that transport entities rom one stock to another at a certain speed, without changing quantity, it will be enough to know values at any end o the pipeline, because on the other side, low value will be the same, but with opposite sign. Keeping in mind that low is representing a rate at which stock value is changing and the thing that ollowing the described approach or turning monitoring results into analytic representation gives us an equation describing stock changes over time, we can get a low value by getting derivative o a stock unction. 2) Stocks monitoring. This case is much more complicated comparing to processing results o monitoring or input and output lows. Key issue is that rom monitoring result we got is a state o the stock in each moment in

5 time during monitoring, but this doesn t give us inormation about inluence o input and output made by lows connected to the stock. The only thing which is known or us is a result o a unction which is a superposition o input and output lows o the stock. In other words, at any moment t 1 value o the stock can be represented as F (t 1 )= InputFlows (t) OutputFlows 1 (t). Order o unctions superposition is depending on model structure, because, according to System Dynamics models speciication, lows execution order depends on their order deinition in the model. Situation with stocks monitoring can became even worse, i a stock has multiple input and output lows, what is common in some cases, when model components are tightly couple to each other. Such situations are oten seen in economical and environment-related models. But, there are also cases, when stocks monitoring works well. This is the case when we are dealing with lea-nodes o the model and these nodes have only one input or output low. In this case stock monitoring can be treated as a input or output low monitoring and we will come to the irst case. 3) Hybrid mode. Hybrid mode means that some o the stocks are monitored directly and others are monitoring on a level o their input and output lows. This approach inherits all advantages and disadvantages o both described approaches, but, it can bring beneits upon identiication o system monitoring strategy. For example, we have a system, which consists o three jars o some liquid. These three containers are connected in chain via two ilters cleaning liquid on the way rom then irst jar to the third one. Dealing with liquid low measurement is non-trivial thing, comparing to measurement o its volume. That is why, or the irst and third jars we can use simple volume (or weight measurement liquid density is known) monitoring or the irst and third jars and perorm lows monitoring only or the second one which is aected by superposition o input and output lows. VII. RESTORATION OF NON-MEASURED STOCKS AND FLOWS In some cases, modelization process can be enhanced via model simpliication process. Interms o this section, under simpliication we assumes not optimization o the model itsel, but incremental cut-o process on a way o model lows incremental deinition, where we are dropping already reconstructed lows and continuing processing only the rest o the model, until we ace a tightly couple model piece that can be processed any urther. Basically, lows are describing a rate at which this particular low changes connected stock value. So i a stock has multiple input and output lows, its value at every moment can be calculated as a sum o all itput lows minus sum o all output lows, like this: F (t) = InputFlows OutputFlows (t) 1 (t) In a corner case, i a stock has only one input or output low, we can easily reconstruct its low by getting a derivative o a unction that represent stock value change over time. In this case such stock and easy reconstruct-able lows can be solved and removed rom the model, to leave smaller structure or urther processing. This recurrent and incremental process ends when there will be no single-low stock let int he system. This procedure can be described with the ollowing algoritmh: 1) Find all unambiguously deined lows by lea stocks. I we know the stock unction and only one low will be connected to it, we also know that this low has an identical inluence on it, and it is possible to deine that F (t) (t). 2) Find all unambiguously deined lea stocks by lows and processed as in the previous step. 3) Simpliication. Analysis all known data and remove inluence o all known lows by a unction subtraction. For example: we have three stocks F, G, H, and two lows (rom F to G) and g (rom F to H). We have measured data o F and G stocks. By the irst step o this algorithm we can deine low inluence unction. We know, that F (t) = (t) g(t) and we can ind the g unction by substitution o rom F. 4) Repeat algorithm while we have any unambiguously deined stocks and lows. At the algorithm stop s we have a simpliied version o an initial system dynamic model, in which we can see only ambiguously lows and socks. Fig. 5 depicts described algorithm. I it is a not complicated model, we can go orward to the next modelization method step. But in the dierent case, the computation complexity may be too high to get modelization done in an acceptable time, and researcher should be to return to experiments and extends sensors environment. VIII. MODELIZATION EXAMPLE To show how this method works on practice let s get back to the initial example shown on ig. 1. Let s assume, that we have a simple production line which takes some item rom an input storage, perorms its initial processing, than, rom initial processing results it is taken or inal processing and ater that it goes to the inal production store. When we started monitoring we know that at the beginning o production cycle we have 8 blank pieces that will be turned into products. In terms o this example blank pieces are converted into product 1:1. When we started production monitoring we got stocks states distribution over time shown in table I. In this table S 1 a stock representing starting warehouse, S 2 is a box with initially processed pieces and S 3 is a stock with production pieces

6 System Dynamics model Model has unambiguos element Yes Find unambiguous stocks and lows Deine unambiguous stocks and lows Simpliy model by removing known lows Simpliication completed Fig. 5. Model incremental simpliication TABLE I. SAMPLE PRODUCTION PIPELINE MONITORING RESULTS No Time S 1 value S 2 value S 3 value From our model deinition we see that stocks S 1 and S 3 are aected by only one stock. It means that we can calculate low value or them directly based on monitoring results. Obviously or this simple case, both approximations will be linear unctions y S 1 = 2x and y S 3 = x ater irst item passes though the model. According to the deinition o low in system dynamics, it is the rate at which the stock is changing at any given instant, they either low into a stock (causing it to increase) or low out o a stock (causing it to decrease). From them math we know that change rate or the unction is unction derivative. In our case, lows will turn into -2 or S 1 out low and +1 or S 3 in low. Other sides o the lows connected to S 2 will have the same values, but with opposite sign, because nothing is disappearing rom the low and nothing new appears. To see some corner cases, let s use a bit more advances example. Assuming that we have one more processing step between stocks S 2 and S 3 and monitoring results appeared like they are shown in the table II Here we can see, that we have two lows which are aected stocks S 2 and S n ew. Monitoring data shows, that everything what get s into the new stock - immediately sent to the S 3. Input low which comes rom S 1 to S 2 is known, it is +2. TABLE II. EXTENDED SAMPLE PRODUCTION PIPELINE MONITORING RESULTS Time S 1 value S 2 value S new value S 3 value Based on this and values o S2 that was measured, we can igure out, that output low rom S 2 is 1. But the same technique will not work or the S new. We expect, that input low is +1, but, we can estimate an output one. at minimum, it is equal to the input low, but we are not able to estimate its maximum. This example gives us two important outcomes: 1) Modelization process do not describe system properties and actual characteristics. It allows to ormalize actual current system behaviour. Complete modelization process may require execution o system under dierent workload to see perormance and boundaries o all lows. 2) During the modelization it is easy to identiy lows where overlow (situations when input lows o a stock exceeds output lows) or underlow (situations when output lows o a stock exceeds input lows, like on the second example). In many cases, detection and addressing o such situations in one o the most important cases during system tuning and optimization. IX. METHOD EVALUATION APPROACH Described method was tested on synthetic data sets and shows its eiciency in achieving stated goals. As a baseline or the experiments was taken a library o test system dynamics test models which is used or evaluation and testing o various computation engines developed and maintained by tools community o System Dynamics society [7]. As a models execution engine was selected engine called PySD [8], because it is developed on a Python language and it appears aster and easier to make code modiications or method evaluation purposes than in other available execution engines. Evaluation and testing o proposed approach was perormed on acilities on sdcloud project [9]. Experiment was perormed on 50 dierent models and was designed in a ollowing way. 1) Basic model execution with advanced logging. on this step we executed 50 dierent test models using modiied execution engine. Goal o the modiication was to log values o all lows beore entering or right ater leaving model stocks. This gives us extended model execution results which contains not only time series describing stocks history, but also input and output lows value or each stock

7 2) Ater execution is completed, rom the model deinition excluded all equations representing lows. This turns a model into pure structure model that contains only rame o the model. 3) All time series describing stocks values rom modelling results are converted into closes analytic representation. According to the approach deinition both approaches were tested: with simple parallel approximation using dierent unctional patterns and selection o best option by estimating approximation quality and with application o trained neural network or time series shape detection or uture approximation using top three detected shapes. 4) Since we had all input and output lows (rom extended model execution logs), or each model we were able to reconstruct its lows via deinition o approximation unction based on logged input and output low values X. METHOD EVALUATION RESULTS As a result o our experiment was ound that or majority o models, 89% o lows across all test models, were recovered correctly. Correctness o the reconstruction o each low was identiy by comparing lows equations in original and newly created model. Flow reconstruction was count as successul i correct low shape was used without taking into account contestant term and allowing slight deviation o coeicients or lower power terms. Parallel approximation shows a bit better results in selection o approximation unction comparing to application o neural network-based classiier, 89% against 75%. But second option shows much better results in perormance and resources consumption. Most probably, this gap can be reduced in case o better training o neural network. Drat implementation o proposed eature was implemented within sdcloud platorm application and integrated with the service or continuous modelling, to leverage its mechanisms or accepting input data streams or urther processing within the platorm. This work was supported by grant IND a rom RFBR. XI. CONCLUSION In this article was described an approach aimed to help building System Dynamics models out o existing system by integrating system monitoring, data analysis and model execution practices. Proposed approach signiicantly depends on abilities or monitoring o target system and has number o limitation, but or many system dynamics application areas it can signiicantly improve modelization process and make it easier and aster than manual system investigation. Proposed method is trying to leverage both classical statistics methodology and modern machine learning mechanisms allowing to solve some issues with the higher run-time perormance and less resources consumption, but, in the same time requiring special eorts on a preparation step, like training o neural networks. REFERENCES [1] Business Dictionary, Web: /deinition/system-dynamics-sd.html [2] Kermack WO, McKendrick AG (August 1, 1927). A Contribution to the Mathematical Theory o Epidemics. Proceedings o the Royal Society A. 115 (772): doi: /rspa [3] Hethcote H (2000). The Mathematics o Inectious Diseases. SIAM Review. 42 (4): [4] Smirnov, V.Y., Kuznetsova, A.V., Approximation o experimental data by solving linear dierence equations with constant coeicients (in particular, by exponentials and exponential cosines), Pattern Recognit. Image Anal. (2017) 27: [5] Belikovich, V.V., Approximation o experimental data by exponential unction, Radiophys Quantum Electron (1993) 36: [6] Lipton Z., Kale D., Elkan C., Wetzel R., Learning to Diagnose with LSTM Recurrent Neural Networks, arxiv: [7] Repository with System Dynamics test models: [8] Python-based execution engine or System Dynamics models PySD: [9] sdcloud - cloud-based platorm or system dynamics modelling: [10] Penskoi A., Gaiosh A., Platunov A., Kluchev A. Specialised computational platorm or system dynamics: ISBN , ISSN , DOI: /sgem2018/2.1, 2 July - 8 July, 2018, Vol. 18, Issue 2.1, pp