Results

For evaluation the Database was initialised with 40 samples, which were randomly drawn from a large repository acquired in the first run of the AOLA-project (cf. Chapter 4). These probes were manually assigned to one of 13 categories (cf. Table 5.3).

Another 20 samples were drawn for testing as listed in Table 5.4. The method chosen for testing was greedy, taking the single best nearest neighbour. Broadening the search by trying on various interpretations will further improve results. This will definitely be the next step to take.

All in all more than 90 percent of the fields were categorised correctly. As you can see in Table 5.1 only 12 of the 100 visible fields were not classified correctly. The visible fields make up 75 percent of the total number of fields classified in the test samples.

Table 5.1: summarised results

	number	misclassified	pct correct
visible fields	100	12	88%
hidden fields	34	3	92%
total	134	15	91%

The results are illustrated detailed in Figure 5.2. Bars representing the individual samples (as listed in Table 5.4) divide into the number of visible fields in the upper part and the number of hidden fields. For a clear differentiation the number of fields that were not classified correctly are shaded. Two main causes for misclassification have been identified. This underlines, that the assignment of the initial probabilities is a very delicate step as well as the importance of a broad database, which offers the possibility to transmit diversified queries.