Artifical neural networks as subsymbolic process descriptors

Hydroinformatics proceeds into that which M.B. Abbott has characterised as the 'post-symbolic' era along two different paths. Along the one path, it elaborâtes tools and even more generai working environments for engineers, environmentalists and other Professionals that make little or no use of symbols in any conventional sense, but which instead work almost entirely with signs. Along the other path, as illustrated in this present work, hydroinformatics increasingly uses non-symbolic, and indeed strictly sub-symbolic, methods in order to construct these tools and more generai working environments. Of course, in this latter case, the constructor of these instruments must still make recourse to symbolic representations, but, as explained here, these are employed essentially as aids to the thinking processes of the constructor, and are not carried over or incorporated in any way into the operations of the constructions themselves. It is this second path of sub-symbolic constructions that forms the subject of the present work. The first chapter of this work is given over to an introduction to the sub-symbolic paradigm in generai and within the particular context of hydroinformatics. The three main current divisions within the sub-symbolic paradigm are those of artificial neural networks (ANNs), evolutionary algorithms and cellular automata. A detailed description of artificial neural networks and, in particular, feed-forward, multi-layer perceptrons, which constitute the sub-symbolic tool used throughout this work, is given in Chapter 2. A brief description of the other two sub-symbolic methodologies is given in Chapter 3. In Chapter 4, ANNs are used to model the rainfall-runoff process using artificially-generated data, laboratory-experimental data and real, measured-catchment data to an exceptionally high degree of accuracy. The problem of extrapolation is described and a possible solution is posed for this problem. In Chapter 5, ANNs are applied to the problem of finding an accurate and stable solution of the pure advection equation. The resulting ANNs provide results that are at least as good as those obtained by more traditional numerical methods. Finally, in Chapter 6, ANNs are applied to the more generai problem of data mining as exemplified by rainfall-runoff modelling, sait intrusion and sediment transportation. It is shown that the A N N outperforms more traditional, essentially manual, methods of data mining, and also provides significantly more accurate results than those obtained by another sub-symbolic paradigm, namely that of genetic programming.