Artifical intelligence applied for forecasting in enterprise decision support
StatSoft Polska nie prowadzi sprzedaży tej książki
|Autor: Tadeusz A. Grzeszczyk|
Wydawca: Wydawnictwo Instytutu Organizacji Systemów Produkcyjnych, Politechnika Warszawska
Wydanie: Warszawa, 2005
Liczba stron: 114
Ogólne informacje o książce
Main subject of the present monograph is to determine forecasting methods to be used in enterprise decision support which would be more effective than – being applied now – single neural networks. The author is convinced that there exists some possibility and need for integration of artificial neural networks and rough set method to be used in decision support system in a company. Such a new method is to be based on analysis of quantitative and qualitative factors.
In practice various quantitative methods are widely spread, based on conventional statistical methods. Very often, the forecasts referred to intensity of anticipated socio-economical situations are also based on opinions of experts representing wide experience and intuition (qualitative methods); however, both the above attitudes do not guarantee reduction of forecasting errors. In some cases, if required, more than one method should be applied, belonging to one or both the aforementioned types of methods. In result, the final forecast is usually determined as simple arithmetic mean or weighted mean of partial forecasts (so-called combined forecast).
Anyhow, even these combined forecasts applied up to now, are not always considered as precise enough. In this connection, with different results, some other methods of forecasting are applied e.g. with the use of artificial neural networks. The latter ones are in many cases better than conventional qualitative and quantitative methods. Detailed checking has been made to prove that in some cases the conventional statistical methods give better results than forecasting methods with the use of single neural networks. All the above discussed methods give satisfactory effectiveness in average conditions, as this is characteristic feature of statistical models in which so-called historical data are used to estimate the relations.
There is no doubt that all the methods mentioned above do not prove true in situations where various non-typical cases take place. For example, in company sales forecasting process, if wide advertising campaign or promotion of some assortment or service is conducted, such situations are considered as special cases. Such non-typical cases are noted relatively seldom in the model process described herein. For single neural networks, all the observations are submitted to averaging operation. In result, the model obtained in this way, is not accurately coordinated with seldom occurring non-typical data. As we know, the neural network is capable to learn even non-typical patterns. However, such learning usually leads to deformation of neural network operation as refers to the patterns which are not included in the learning process.
The main purpose of the present monograph has been to solve the research problem that called for determination of short-term forecasting method in decision support system in a company, which would be more reliable – i.e. with smaller forecasting error – than prediction based on single neural networks.
In the present monograph, a new concept of forecasting system has been formulated. It comprises the following three main sub-systems:
- quantitative analysis with application of neural network,
- qualitative analysis based on rough set theory,
- sub-system integrating the results derived from both the above sub-systems.
The research process was performed in three planes. The first one consisted in research on neural networks in short-term, time series analysis. The second one was referred to application of rough set theory for qualitative analysis of problems connected with non-typical situations analyzed in such cases as e.g. promotion or advertising campaign. The third research plane was related to implementation and research of forecasting system integrated (coupled) in a parallel way.
Apart from neural networks and the method of approximate conclusions, in experimental part of the present monograph also the other tool belonging to the group of methods connected with artificial intelligence, so-called genetic algorithm, was applied. The algorithm was used during creation of neural models, for selection of values, arranged in time series prior to prediction. Such configuration allowed for preliminary determination of number of neurons in input layers of the networks subjected to research.
Implementation of qualitative subsystem with application of rough set theory was performed with the use of Delphi programming system in its 5.0 version. Qualitative forecast supporting system proved helpful for empiric verification of research assumptions. Presentation of ready-made practically applicable solution in the present monograph created effective basis for further works on coupled forecasting method to be used for instance in practical business applications.
Empiric verification of particular methods in programming process was carried out and based on real data collected in large commercial company. These data were arranged in time series describing daily sales volume of one product within about one year.
The author made research on forecasting models built with the use of four types of networks: Linear, MLP (Multilayer Perceptron), RBF (Radial Basis Function), GRNN (Generalized Regression Neural Networks). In order to compare the results obtained with the aid of models built on the basis of single neural networks MLP, with those received at the output of coupled forecasting system, the author applied (apart from quality type specifications in the form of diagrams) the quantitative parameters to ensure accurate and objective confrontation of results. The following quantitative measures of quality of neural models were applied for the purpose of present monograph (among the others): regression statistics, mean squares error, average relative errors, one- and five-day forecasts. Simulated forecasts, determined with the aid of integrated prediction forecasting system, showed significantly smaller errors when compared with the errors calculated with use of single neural networks. The results of experiments confirmed the thesis assumed by the author that there exists possibility as well as requirement for integration of coupled neural networks with rough set method for forecasting in enterprise decision support. Anyhow, all the above results encourage further research and study on various methods to be applied in support decision system in a company.
Basing on formulated in points 2.1.-2.4., general methodical guidelines of study, in further considerations the author presents the forecasting methods applied so far (points 2.5.-2.8.). This analysis of state of knowledge in discipline of methods of prediction in enterprise includes also (comparatively well-known and described in spacious literature) the principles of constructing neural models of time series, used in further part of study. In further course (in chapter third), considerations were concentrated on problem of using within the process of expectations, the qualitative methods of rough set. That created a starting to presentation implementation of complex forecasting method (point 4.1.). Results of verification of models based on single neural networks as well as research of complex forecasting method are presented in points 4.2.-4.4. In recapitulation (chapter 5.) a position on achieved theoretical and applicant purposes of study was taken. There were also defined the intentions concerning the further research studies of the author.
2. Forecasting Methods
2.2. Purpose, Subject and Range of Research
2.3. Economic Reason for Undertaking the Subject
2.4. Research Programme
2.5. Qualitative Methods
2.6. Quantitative Methods
2.6.1. Time Series
2.7. Instruments of Artificial Intelligence
2.8. Neural Networks in Forecasting of Time Series
2.8.1. Characteristics of Neural Models
2.8.2. Principles of Designing Time Series Neural Network Model
2.8.3. Evaluation Methods for Neural Models
3. Use of Rough Set Theory in Qualitative Prediction Subsystem
3.1. Formulation of Problem to be Solved
3.2. Plan of Research Connected With Qualitative Subsystem
3.3. Information System
3.4. Demonstration of Sample Information System
3.5. Rough Concepts
3.6. Definitions Related to Reduction of Superfluous Attributes
3.7. Example of Determining and Investigations of Decision System
3.8. Algorithm for Discovering Decision Rules
3.9. Results of Experiments – Decision Rules
3.10. Selection of the Method of Integration: Both Qualitative and Quantitative Subsystems
4. Research Referred to Intelligent Systems
4.1. Implementation of Integrated Forecasting Method
4.1.1. Purpose and Guidelines of Implementation
4.1.2. Tools for Implementation
4.1.3. Description of Prepared Software
4.1.4. Conclusions Relating to Implementation
4.2. Data Used in Experiments
4.3. Neural Network Time Series Forecasting
4.3.1. Use of Genetic Algorithm for Calculation Related to Number of Neurons in Input Layer
4.3.2. Preparation of Models for Verification of Effectiveness of Single Neural Networks
184.108.40.206. Investigation of Linear Models
220.127.116.11. Preparation of RBF Network
18.104.22.168. Experiments with GRNN Networks
22.214.171.124. Tests with Perceptrons
126.96.36.199. Application of Serial-Coupled Model
4.3.3. Empirical Verification of Single Neural Networks
4.3.4. Selection of a Model for Quantitative Analysis Subsystem
4.4. Integration of Quantitative and Qualitative Results
4.5. Conclusions Related to Research
5. Summary and Final Conclusions
5.1. Theoretical Problems and Effects
5.2. Empirical Research
5.2.1. Neural Networks
5.2.2. Rough Set Method
5.2.3. Coupled Forecasting Method
5.3. Directions of Further Research
List of Figures
List of Tables