Microsoft Word - brain_vol8_issue3_v6_ok1.docx 69 Artificial Intelligence as a Decision-making Tool in Planning the Research Simina Maris Ioan Slavici University, Prof. Dr. Aurel Păunescu Podeanu, Timișoara, Romania Politehnica University of Timisoara, Piața Victoriei 2, Timișoara 300006, Romania Phone: 0256 403 000 siminamaris@gmail.com Titus Slavici Politehnica University of Timisoara, Piața Victoriei 2, Timișoara 300006, Romania Phone: 0256 403 000 titusslavici@yahoo.com Petre Nenu Politehnica University of Timisoara, Piața Victoriei 2, Timișoara 300006, Romania Phone: 0256 403 000 nenu_florinel@yahoo.com Liliana Baciu Politehnica University of Timisoara, Piața Victoriei 2, Timișoara 300006, Romania Phone: 0256 403 000 lforgacs@bleu-electrique.ro Abstract The quality of a finite product is influenced both on quantitative and qualitative factors and thus it is somewhat difficult to determine the major factors which affect it and their degree of influence. In this paper we present the usage of artificial intelligence (in particular artificial neural networks) in the development of an efficient research plan for studying the quality of finite products - in particular, wood briquettes obtained from various biomass mixtures. Keywords: Artificial intelligence, Neural networks, Innovation, Finite products 1. Introduction There is a strong correlation between basic research and applied research. The linear process linking basic research to applied research, development, commercialization, diffusion and their consequences is commonly denoted as technological innovation and development (Zahra et al, 2007). During the last century, innovation became gradually the most important factor leading to productivity growth. Consequently, one cannot omit the role of academic research and its applications in the socio-cultural and economic development. Thus, new strategies to create wealth by valorizing existent results of academic research were developed. Once agreed on the importance of research (both basic and applied) for technological development and, ultimately, for gaining wealth, it becomes simple for a company to search for new solutions and improvements regarding their products in the academic environment. Thereby, real- life processes and phenomena are modeled and the subsequent physical, chemical or mathematical models are studied for their properties. Development of computers and computer science lead to new methods to analyze a model. While methods of numerical analysis existed since more than 2000 years, the speed performances of modern computers lead to numerical solutions of increasingly complex problems, from numerical BRAIN: Broad Research in Artificial Intelligence and Neuroscience Volume 8, Issue 3, September 2017, ISSN 2067-3957 (online), ISSN 2068-0473 (print) 70 weather prediction to programming automatic pilots and even chemical properties of semiconductors. All the models mentioned above use quantitative input values in order to compute quantitative, numerical outputs. A problem arises when, during the modeling phase, qualitative (discreet) inputs are required. The classical models should be replaced by new, statistical models, created either by statistical methods or by using artificial intelligence tools. Our work is a response to the need of innovation expressed by a private company, SC Andrei Slavici SRL. The company is searching for innovation in the production of wood pellets and briquettes on the market of nonconventional bio-fuels. There is a positive global trend in the usage of renewable sources of energy and Romania has already overcome the 22% threshold of renewable energy used, assumed for 2020 (REN 2015). In this regard, economic, environmental and political aspects of using biomass in energy are investigated worldwide (Davis 2014, Nishiguchi 2016), and especially the usage of wood pellets as solid biofuel (Grammelis 2010). 2. Methods Every research activity contains a data collection phase. Data collection ensures there is enough real life information to construct a model and / or to set its limits. Data may refer to (Maris 2013, Greer 2015, Untaru 2012): • physical or chemical properties of an object (such as density, thermal conductivity, melting point, etc.) • biological models (e.g., “artificial” neurons used to model “real” neurons) • initial values for a process when dealing with time-progressive models (such as initial temperature, initial concentration, etc.) • boundary values for an object or process (such as temperature on the edges of the studied object, velocity of the edges, etc.) • economic information (prices, productivity, demand and offer, etc.) • miscellanea (state of economy, amount of rain, client’s degree of satisfaction, provenience of the object, availability of the object in a specific area, etc.) The traditional statistical methods used to link a certain output variable given a set of input conditions are generically denoted by the term of “curve fitting”. Among them, one of the most commonly used statistical methods is the least squares method. New techniques of forecasting were developed as consequence to the development of computers and Artificial Intelligence. Instead of “curve fitting”, the term “pattern recognition” appeared, designating a process through which a label is assigned to a given input value. Among the pattern recognition software currently used in the field of decision making, there are (knowledge-based) expert systems and artificial neural networks (ANNs). Expert systems emulate a real expert. Hence, the output obtained using an expert system is comparable to a decision of a most competent real expert (Corvid Exsys). However, no decision could be made in the absence of the knowledge base. Usually, the knowledge base is a collection of rules with the general form “IF condition THEN result.” The process of decision making is based on the rules previously loaded in the knowledge base, leading to an answer to the problem (question) in discussion. ANNs became widely used since 1980s. The “neurons” of an ANN are software elements which emulate human neurons by processing some input data and returning a function of it. Currently, ANNs are considered as one of the most accurate decision-making tools. ANNs overcome the restrictions of classical methods used for complex problem solving like forecasting electricity consumption (Kaytez et al., 2015), forecasting wind power and control the parameters of wind turbines (Ata, 2015), analysis of consumer choices (Kennedy at al., 2016) and estimating gross calorific values (Estiati et al., 2016). To use an ANN, one should go through three phases (Russell and Norvig 2002): S. Maris, T. Slavici, P. Nenu, L. Baciu - Artificial Intelligence as a Decision-making Tool in Planning the Research 71 • the training phase (during which known associations between inputs and outputs are fed into the network); • the validation phase (during which the outputs of the ANN are validated against known values of the outputs); • the testing phase (in which the ANN predicts outputs associated to a certain set of inputs). Since ANNs are generally more accurate than expert systems, linking both quantitative and qualitative inputs in order to obtain an output, we will use ANNs to achieve our goals. A typical ANN contains (Figure 1): • an input layer – I, composed of a certain number of neurons, • one or more hidden layers – H (each composed of a certain number of neurons) • an output layer – O, usually composed of one neuron (the “answer” of the network). Figure 1. Example of multilayer neural network (source: Samarasinghe, 2016) The computations were made using the Neural Networks toolbox of Matlab (Chapman, 2015). 3. Case study Our research started with a request from the company SC Andrei Slavici SRL to help innovate their production. A private microenterprise, SC Andrei Slavici SRL was founded in 2012 as a spin-out of the Ioan Slavici Foundation for Culture and Education. Its purpose is to capitalize commercially the results of scientific research. The NACE code of the company is 1629 (Manufacture of other products of wood; manufacture of articles of cork, straw and plaiting materials), which allows the company to produce fire logs (briquettes) and fire pellets out of pressed wood or substitute materials. The technological line of the company allows the processing of a minimum of 400 kg raw material at once. The company search innovation in order to produce fire pellets and briquettes with higher energetic efficiency, addressing thus to the growing market of nonconventional bio-fuels. A more detailed description of the situation was given by Maris et al. (2016). The key process parameters of pelletization have been listed by Stelte in 2014: wood species (chemical composition), particle size, content of moisture, conditioning temperature, pelletizer temperature, length and diameter of the die, rotation speed of the die and, in case of torrefied biomass, the degree of torrefaction. Other important aspects of the problem are the chemical composition of the resulting ash (Biedermann and Obernberger, 2005), as it can damage the stoves, and ash melting temperature, a higher temperature meaning a more efficient combustion (Holubick et al, 2015). In addition to these aspects, the price of raw material should not be neglected, as it affects the price of the finite product and hence its attractiveness to customers. The price of raw BRAIN: Broad Research in Artificial Intelligence and Neuroscience Volume 8, Issue 3, September 2017, ISSN 2067-3957 (online), ISSN 2068-0473 (print) 72 material depends on demand/offer ratio, frequency of cataclysms, state of economy, etc. (Untaru et al, 2012) For the moment, the chemical composition of certain wood types will be considered. The chemical composition is known (Krajnc, 2015) and will be used further in the computations. Using the Mendeleev formula, the total calorific value can be computed for various combinations of wood. An example of these values can be found in Table 1. However, while the Mendeleev formula links the calorific value of the fuel to only 4 chemical components, we aim to feed the neural network with an extended chemical composition of the fuel (10 chemical components). Table 1. Calorific values for various types of wood and combinations of woods (sample) Biofuel C H O S Calorific value (Mendeleev formula) unit (%,wt. cont.) (%,wt. cont.) (%,wt. cont.) (%,wt. cont.) (MJ/kg) Bark 51.400 5.700 38.700 0.085 19.081 Beech 47.900 6.200 43.300 0.015 17.693 Poplar 47.500 6.200 44.100 0.031 17.679 Spruce 49.800 6.300 43.200 0.015 18.658 Willow 47.100 6.100 44.200 0.045 17.420 Beech and Spruce 1:1 48.850 6.250 43.250 0.015 18.279 Spruce and Beech 2:1 48.850 6.250 43.250 0.015 18.405 Beech and Spruce 2:1 48.850 6.250 43.250 0.015 18.152 Figure 2. ANN structure and performance S. Maris, T. Slavici, P. Nenu, L. Baciu - Artificial Intelligence as a Decision-making Tool in Planning the Research 73 For our purpose, we used 27 samples, each sample containing 10 different elements, referring to unmixed wood. The set was divided randomly in training set (70% of the sets of values, i.e. 19 sets), validation set (15% of the sets of values, corresponding to 4 sets) and testing set (15% of the sets of values, corresponding to 4 sets). The data division was performed randomly. The best fitting results were obtained when training the ANN using Levenberg-Marquardt algorithm, corresponding to an overall accuracy of 99.57%. Figure 3. Regression plots for the case presented Table 2. Comparison of the calorific values obtained by using the Mendeleev formula and the ANN (sample) Biofuel Real calorific value (Hartmann et al, 2013) Calorific value Mendeleev formula (4 chemical components) Calorific value ANN (10 chemical components) unit (MJ/kg) (MJ/kg) (MJ/kg) Bark 19.200 19.081 18.9322 Beech 18.400 17.693 17.4034 Poplar 18.500 17.679 17.5016 Spruce 18.800 18.658 18.5133 Willow 18.400 17.420 17.2842 Beech and Spruce 1:1 - 18.279 17.9313 Spruce and Beech 2:1 - 18.405 18.1186 Beech and Spruce 2:1 - 18.152 17.7503 BRAIN: Broad Research in Artificial Intelligence and Neuroscience Volume 8, Issue 3, September 2017, ISSN 2067-3957 (online), ISSN 2068-0473 (print) 74 The calorific value computed by the ANN (based on 10 inputs) is generally less than the calorific value computed by Mendeleev formula (based on 4 inputs). As further work, the inputs will consider the rest of the influencing parameters, and the outputs of the neural network will also contain the amount of noxious by-products and the price of the finite product. The performance of an ANN is influenced both by its architecture and the algorithms used during the computations. While it is important to find the best combination of architecture and algorithms, it is not practical to perform this experiment even using the newest computation resources. Thus, in order to optimize the ANN, one should determine first the network parameters with the greatest influence on the forecast accuracy. Once these parameters are determined, statistical methods should be used to construct a response surface in order to determine the best values for the influencing parameters. (Slavici 2006) However, as this represents work in progress, these results should be presented in another paper. The results obtained using ANNs are useful to discern which inputs (and with what weights) determine the desired outputs. This is useful because it will help to perform the required practical experiments with reduced costs. 4. Conclusions Artificial Intelligence provides the researcher with versatile tools, which can be successfully used in the process of planning the research. ANNs can be successfully used to estimate the yield of a certain combination of inputs in a given situation, despite some practical problems such as: a high complexity of the mathematical model for the required situation, associated with an insufficient accuracy of this model; incomplete available data, affected by noise and bias; too many restrictions to be applied and simultaneously optimized in the process. Moreover, they allow the researcher to process both quantitative and qualitative data. Currently, we constructed an ANN to process the inputs associated to the fabrication of wood pellets and briquettes. As input, it was considered only the chemical composition of the biomass, the output being the net calorific value. The overall accuracy of the network was 99.57%. As a further work, the inputs will be diversified and the ANN will be optimized, thus leading to more accurate forecasts. References Ata, R. (2015). Artificial neural networks applications in wind energy systems: a review. Renewable and Sustainable Energy Reviews, 49, 534-562. Biedermann, F., & Obernberger, I. (2005). Ash-related problems during biomass combustion and possibilities for a sustainable ash utilisation. Austrian Bioenergy Centre GmbH & BIOS Bioenergiesysteme, Graz, Austria. Retrieved June 25, 2016, from http://www. biosbioenergy. at/uploads/media/Paper-Biedermann-AshRelated-2005-10-11.pdf Chapman, S. J. (2015). MATLAB programming for engineers. Nelson Education. Corvid Exsys. (2015). Retrieved from http://www.exsys.com/exsyscorvid.html Davis, S. C., Hay, W., & Pierce, J. (2014). Biomass in the energy industry: an introduction. London, United Kingdom: BP plc. Estiati, I., Freire, F. B., Freire, J. T., Aguado, R., & Olazar, M. (2016). Fitting performance of artificial neural networks and empirical correlations to estimate higher heating values of biomass. Fuel, 180, 377-383. Grammelis, P. (2010). Solid Biofuels for Energy: A Lower Greenhouse Gas Alternative. Springer Science & Business Media. Greer, K. (2015). New Ideas for Brain Modelling, Broad Research in Artificial Intelligence and Neuroscience, vol. 6, no. 3-4, pp. 26-46. S. Maris, T. Slavici, P. Nenu, L. Baciu - Artificial Intelligence as a Decision-making Tool in Planning the Research 75 Hartmann, H., Reisinger, K., Thuneke, K., Höldrich, A., & Roßmann, P. (2013). Handbuch Bioenergie-Kleinanlagen, ed 3. Überarbeitete Auflage. Fachagentur Nachwachsende Rohstoffe eV, Hartmann, H. Holubcik, M., Jandacka, J., & Malcho, M. (2015). Ash melting temperature prediction from chemical composition of biomass ash. The Holistic Approach to Environment, 5(3), 119-125. Kaytez, F., Taplamacioglu, M. C., Cam, E., & Hardalac, F. (2015). Forecasting electricity consumption: a comparison of regression analysis, neural networks and least squares support vector machines. International Journal of Electrical Power & Energy Systems, 67, 431-438. Kennedy, M., Dinh, V. N., & Basu, B. (2016). Analysis of consumer choice for low-carbon technologies by using neural networks. Journal of Cleaner Production, 112, 3402-3412. Krajnc, N. (2015). Wood Fuels Handbook. Food and Agriculture Organization of United Nations. Pristina, 2015. Maris, S. (2013). Numerical characterization of the computational domains for crystals grown by Bridgman technique. In Recent Advances in Low-temperature Processes for the Development of RF, EO, Magnetic and Electronic Bulk and Thin Film Crystals. Materials Science and Technology (MS&T), October 27-31, 2013, pp. 2768-2773. Montreal, Quebec, Canada. Maris, S., et al. (2016). Financing Methods For Innovative Start-Ups In The Domain Of Nonconventional Energies, Nonconventional Technologies Review, No. 3 (XX) / 2016, pp.65- 70, ISSN 2359 – 8654. Nishiguchi, S. & Tabata, T. (2016). Assessment of social, economic, and environmental aspects of woody biomass energy utilization: Direct burning and wood pellets. Renewable and Sustainable Energy Reviews, 57, 1279-1286. Ren, P. S. (2015). Renewables 2015 global status report. REN21 Secretariat: Paris, France. Russell, S. & Norvig, P. (2002). Artificial Intelligence: A Modern Approach. 2nd Edition. Prentice Hall Series in Artificial Intelligence. Fort Collins, CO. Samarasinghe, S. (2016). Neural networks for applied sciences and engineering: from fundamentals to complex pattern recognition. CRC Press. Slavici, T. (2006). Artificial Intelligence. Eurostampa Publishing House. Timisoara, Romania. Stelte, W. (2014). Optimization of product specific processing parameters for the production of fuel pellets from torrefied biomass. Taastrup, Denmark: Center for Biomass and Biorefinery, Danish Technological Institute. Untaru, M., Rotarescu, V., & Dorneanu, L. (2012). Artificial Neural Networks for Sustainable Agribusiness: A Case Study of Five Energetic Crops. Revista Agrociencia, 46(5), 507–518. Zahra, S. A., Van de Velde, E., & Larraneta, B. (2007). Knowledge conversion capability and the performance of corporate and university spinoffs. Industrial and Corporate Change, 16(4), 569–608. Simina Maris is PhD student with the Politehnica University of Timisoara since 2015. She holds a BSc in Mathemathics and Physics and a BSc in Business Administration, as well as a MSc in Applied Computer Sciences. She is part-time assistant professor at the Ioan Slavici University and head of Mathematics and Sciences Department at the Ioan Slavici High School. Her interests are: applied mathematics, mathematical modeling, practical aspects of using ICT in education, computer simulation, artificial intelligence, using artificial intelligence for optimization of processes. Titus Slavici is Professor with Politehnica University of Timisoara, and also co- founder and professor with Ioan Slavici University of Timisoara. He helds a PhD in Industrial Engineering and a PhD in Finances. Since 2015 he is habilitated in Industrial Engineering. Titus Slavici has obtained a number of R&D grants and succesfully coordinated their implementation. His main areas of interest are: computer science, artificial intelligence, industrial engineering, finances and management, using artificial intelligence for optimization of processes. BRAIN: Broad Research in Artificial Intelligence and Neuroscience Volume 8, Issue 3, September 2017, ISSN 2067-3957 (online), ISSN 2068-0473 (print) 76 Petre Nenu is PhD student with the Politehnica Universityof Timisoara since 2015. He helds a BSc in Heat Machinery and Equipments and a MSc in Renewable Energies. Currently, he is engineer at Colterm SA. His interests are related to heat production, thermodynamics, applied sciences, nonconventional energy resources, sustainable agricultural development, sustainable agricultural production, applied computer sciences, industrial engineering. Liliana Baciu (Forgacs) is PhD student with the Politehnica University of Timisoara since 2016. She holds a BSc in Mechanical Engineering and a MSc in Industrial Engineering. Currently, she is manager at Bleu Electrique Romania. Her interests are: mechanical engineering, chemical technological equipments, welding, industrial engineering, management, using artificial intelligence for optimization of processes.