NEURO-FUZZY MAPPING SYSTEMS FOR REAL ESTATE ANALYSIS AND FORECASTING

Abstract

The motivation for this thesis has been the author's belief that there is no real lack of knowledge in this world. There is, however, a need for greater intersection between pools of knowledge. Firstly, this ensures that the benefits are more likely to outweigh the cost of research because of a wider scope of applicability. Secondly, it reduces the phenomenon of reinventing the wheel, and results in more value-added production and economic activity. Where certain disciplines have remained insular, they become susceptible to the in-breeding of ideas and practices. Hence, they may exhibit thinking that is outmoded and unproductive. Disciplines that have gone the opposite way to merge with other disciplines usually produce genetically superior products. The field of medicine is a prime example of where the marriage of disciplines has produced favourable results. Knowledge of physics and engineering has been successfully integrated into medical diagnoses and treatment. It is in this spirit that the author believes that the benefits of recent developments in the field of Artificial Neural Networks and Fuzzy Logic can be profitably integrated into real estate analysis and forecasting. The aim of the thesis is to demonstrate the integration of neural network and fuzzy logic technologies in simulating decision-making in real estate analysis and forecasting. This aim is achieved via three objectives. The first objective is to build a non-parametric adaptive property mass-appraisal system based on neural network technology. The second objective is to build a qualitative diagnostic system for assessing industry maturity based on fuzzy logic technology. And the last objective is to build a qualitative and adaptive forecasting system based on both neural network and fuzzy logic technologies. To address Objective l, this thesis will introduce the background of Artificial Neural Networks, and explain the architecture for the popular Backpropagation neural network. Subsequently, the Backpropagation network is applied to the valuation of private residential apartments in Singapore. The network is trained using 833 transactions, and the learned connection-weights are tested on an additional 222 transactions. The advantages of the Backpropagation network versus traditional regression analysis are discussed. The Backpropagation network can be used in several ways, namely, as a tool to speed the normal valuation process, for checking discrepancies, and for preliminary valuations before site inspection. Although the Backpropagation network has been traditionally considered a black-box model, this thesis demonstrates a method by which the connection-weights in a BP model may be interpreted for the purpose of causal analysis. Thus important factors contributing to the sale price of an apartment may be ranked using the connection-weights. To address Objective 2, the author has developed the Fuzzy Industry Maturity Grid for diagnosing the maturity of an economic sector or industry. The Fuzzy Industry Maturity Grid is an extension of the traditional Industry Maturity Grid Diagnosis is to identify fast growing sectors and to map out growth strategies for sectors showing signs of maturity. By incorporating fuzzy set theory and fuzzy aggregation models in decision-making, the conventional Industry Maturity Grid is enhanced from a descriptive analysis to a semi-quantitative method which captures an expert's knowledge of the sector or industry. This is done via a linguistic scale describing the characteristics under each of the three main dimensions of the IMG to form fuzzy sets. Subsequently, a hierarchical aggregation of information based on fuzzy aggregation operators is carried out, and a conceptual hypercube is used to determine the rank and ranking size of prescribed strategies. The application of the Fuzzy Industry Maturity Grid is illustrated with an example on the Singapore property sector. To address Objective 3, the author has developed the Adaptive Fuzzy Event Propagation approach to the qualitative modelling of adaptive dynamic systems. Traditionally, system modelling approaches have been quantitative in nature. These quantitative approaches are S commonplace in the realm of the physical sciences. Such models include multiple regression and econometric models. The need for adequate quantitative data can sometimes hinder their employment due to a lack of data; and it may also gives the false impression of modelling precision due to the use of crisp numbers. The proposed Adaptive Fuzzy Event Propagation does not use crisp measured data but instead uses fuzzy numbers to represent expert(s) opinions/interpretations of events. Fuzzy arithmetic operations are perf01med on the fuzzy numbers in an auto-associative neural network. An auto-associative network enables the representation of feedback systems. The Adaptive Fuzzy Event Propagation approach can be used to represent either the opinions of one industry expert, or it can synthesise the conceptual models of several experts into one implementation model. Propagation of the events modelled result in a linguistic forecast of the events. Furthermore, the adaptive dynamics of a system are captured through the use of unsupervised learning neural network algorithm. The Adaptive Fuzzy Event Propagation approach is illustrated on a simple hypothetical system, and the results discussed. The validation of the proposed approach, its advantages and limitations are also discussed.