Keywords: nearest neighbour algorithm, coefficient of volume compressibility, geotechnical parameters, data mining, database, parameter estimation.

The limitation of settlement of a civil engineering structure is a paramount feature of the foundation design process. A knowledge of the underlying geotechnical parameters is therefore critical to a stable design life. As well as the foundation and soil type, structural mass and pressure, foundation settlement is variously dependant on properties of the foundation strata including: elasticity, permeability, coefficient of volume compressibility, coefficient of consolidation and compression index. This work concentrates on saturated engineering soils, namely clays and silts, as a foundation medium - soils exhibiting a permeability that causes full settlement to be delayed as water is slowly expelled from the soil matrix under foundation pressure. Approximate, time independent calculations can make use of a preliminary knowledge of the coefficient of volume compressibility m_v, the average applied pressure, and the thickness of the compressible strata to arrive at a final compression value. Preliminary site investigations would traditionally encompass classification and index testing of the soils using relatively straightforward and inexpensive tests. A data-based approach would attempt to establish a link between the index properties of a database and the compression parameters, using a multi-dimensional association. Subsequently, the association can be extrapolated in some way to predict values of mv from just the index properties. This paper describes the use of nearest neighbour algorithms to generate the link between a number of parameters and m_v in a database and their use in the prediction of the coefficient of volume compressibility.

Nearest neighbour algorithms work by searching a database for records that are similar to a user's record over a number of records. Although the user's record would have one or a number of parameters missing, the nearest neighbour algorithm works on the available parameters so that the similar database records can be utilised to make predictions about the user's missing parameters. Central to the process is a similarity function that finds the distance of the user's record to a database record and is distinctive for a particular nearest neighbour algorithm which will in turn select one or more records from the database with the shortest distance to the user's record to be incorporated into the prediction process.

Four nearest neighbour algorithms were tested together with a linear regression algorithm and a neural network multilayer perceptron. Experiments were carried out on a database containing nearly 180 records, each with an mv value although the database contained 17% missing values overall. Algorithms made associations between parameters in the database prior to excluding database records to enable predictions of missing values to be made.

Overall, results showed that predicted values of the coefficient of compressibility exhibited a high degree of variability. The effectiveness of each of the algorithms was estimated by measuring the error rate of each prediction. This was done by comparing the actual database mv value with the predicted value. Error rates varied greatly from zero to 1300% for outliers using one of the algorithms. The trimmed mean error rates were: IBk 106%, multilayer perceptron 99%, linear regression 96%, LWL 76%, kStar 73%, NnLr 57%. These should be compared with estimations based on commonly used descriptions of the soil type that are given in a range of parameter values where, for example, a firm clay of mv 0.1 to 0.3MN/m² indicates a 100% variation range from the mean and similarly for highly compressible clays the range is 200%. Results showed that nearest neighbour algorithms were better than either linear regression or a neural network algorithm at predicting m_v values from a database, and that the gBase algorithm had the lowest error rate.

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £62 +P&P)