Academic journal article Journal of Real Estate Literature

Hybridizing Cuckoo Search with Levenberg-Marquardt Algorithms in Optimization and Training of Anns for Mass Appraisal of Properties

Academic journal article Journal of Real Estate Literature

Hybridizing Cuckoo Search with Levenberg-Marquardt Algorithms in Optimization and Training of Anns for Mass Appraisal of Properties

Article excerpt

(ProQuest: ... denotes formulae omitted.)

In recent times, property appraisals/valuers in developing economies are getting involved with the mass appraisal of properties. This is precipitated by a rise in property development and investment; increase in mortgage requirements for securities and revaluation by financial institutions; the need to ascertain market value of tax base and subsequent ad valorem purposes by local authorities; and dissolution of company or marriage contracted in the community of property for assets sharing and expropriation of land and properties by state or regional governments. Several appraisal/valuation models are utilized to estimate the market values of properties. To achieve accuracy, valuers must know the suitability of model(s) towards solving an appraisal/valuation problem. Should the valuer rely on single appraisal /valuation models (in this study, these are traditional valuation models) to estimate value, the ability to cope with analysis and interpretation over a number of properties and achieve accuracy is a matter of utmost concern (Adair and McGreal, 1988a). In practice, models that are suitable for mass appraisal/valuation of properties include multiple regression analysis (MRA), additive nonparametric regression (ANR), artificial neural networks (ANNs), expert systems (ESs), fuzzy logic (FL), genetic algorithm (GA), and spatial analysis (SA). Although these models can be used for the appraisal of a single property, their benefits are profound when utilized for the mass appraisal of properties (McCluskey et al., 1997).

Interestingly, for more than three decades, these models have been utilized for the mass appraisal of properties in developed economies. But acceptability of these models has not been without resistance. For instance, Adair and McGreal (1988b), Rayburn and Tosh (1995), and Baen and Guttery (1997) report fears exhibited by appraisers/valuers in the 1980s and 1990s that mass appraisal models may replace them or undermine their power of subjectivity. Mooya (2011) drew a lesson from the impact of technology on the land surveying profession and expressed reservation that a similar fate might affect the property appraisal/valuation industry should these models be embedded into its practice. Plausible as the argument might be, technologies are designed to complement the professional role of a land surveying expert. Similarly, in mass appraisal, these models are designed to complement the roles of human valuers. In effect, certain potential roles of human valuers such as data gathering, assembling and inputting data into computer software, interpretation of results and measurement of errors are essential for smooth operation of these models.

But while using these models, obvious limitations were found in some of them, necessitating hybridization to achieve accuracy. For instance, ES is a versatile model that works under set rules by an expert. Its shortcoming is that it does not learn but operates under set conditions; also if there is a reaction in market operations, the model is only aware of its ''rules,'' thereby reporting once the target (rule) is attained. Additionally, if the process of handling a task is not completely specified, ES will provide an incomplete solution. With this situation, incorporating it with learning algorithms such as BP or GA enhances the results. Again, combining ANN with SA will help detect additional neighborhood(s) that would have been excluded from analysis. Fuzzy logic is designed to deal with vagueness in property attributes, but Zurada, Levitan, and Guan (2011) view its shortcoming to determine of fuzzy sets and rules. Therefore, hybridizing FL with learning tools such as ANN or GA will help mitigate its limitations (Cordón, Harrera, Hoffmann, and Magdalena, 2001).

The concern of this study, however, is ANN and how a combination of algorithms could influence its training for effectiveness. There are many types of ANN: back propagation (BP) feed-forward (multi-layer perceptron) neural networks, radial basis function neural network, Kohonen self-organizing neural network, learning vector quantization neural network, and recurrent neural network. …

Search by... Author
Show... All Results Primary Sources Peer-reviewed

Oops!

An unknown error has occurred. Please click the button below to reload the page. If the problem persists, please try again in a little while.