"... In PAGE 25: ... To maximize lnP (DjN), we need to know the distribution that the network apos;s outputs are to learn. Table1 shows three probability distributions that are commonly assumed in practice, and their appropriate error functions (Rumelhart et al., 1995).... In PAGE 88: ... Table1 0: Test-set error from a ten-fold cross validation. Table (a) shows the results from running three learners without the domain theory; Table (b) shows the results of running three learners with the domain theory.... In PAGE 88: ...5% 34.7% (b) Generating Non-KNN Ensembles Table1 0a presents the results from the case where the learners randomly create the topology of their networks (i.... In PAGE 88: ...opology of their networks (i.e., they do not use the domain theory). Table1 0a apos;s rst row, best-network, results from a single-layer neural network where, for each fold, I trained 20 networks containing between 0 and 100 (uniformly) hidden nodes and used a validation set to choose the best network. The next row, bagging, contains the results of applying the bagging algorithm to standard, single-hidden-layer networks, where the number of hidden nodes is randomly set between 0 and 100 for each network.... In PAGE 89: ...can produce large alterations in the predictions of the networks, thereby leading to an e ective ensemble. The bottom row of Table1 0a, Addemup, contains the results of a run of Addemup where its initial population (of size 20) is randomly generated using Regent apos;s method for creating networks when no domain theory is present (refer to Section 4.3.... In PAGE 89: ... Again, each ensemble con- tains 20 networks. The rst row of Table1 0b contains the generalization results of the Kbann algorithm, while the next row, Kbann-bagging, contains the results of the en- semble where each individual network in the ensemble is the Kbann network trained on a di erent partition of the training set. Even though each of these networks start with the same topology and \large quot; initial weight settings (i.... In PAGE 89: ...e., bagging in Table1 0a). The next two rows result from the Regent algorithm.... In PAGE 89: ... The rst row, Regent-best- network, contains the results from the single best network output by Regent, while the next row, Regent-combined, contains the results of simply combining, using Equa- tion 21, the networks in Regent apos;s nal population. Last chapter, I showed the e ective- ness of Regent-best-network, and comparing it with the results in Table1... In PAGE 90: ... Notice that simply combining the networks of Regent apos;s nal population (Regent-combined) decreases the test-set error over the single-best network picked by Regent. Addemup, the nal row of Table1 0b, mainly di ers from Regent-combined in two ways: (a) its tness function (i.e.... In PAGE 90: ...e., a repeat of Addemup from Table1 0b). The results show that, while reweighting the examples during training usually helps, Addemup gets most of its generalization power from its tness function.... In PAGE 91: ... Table1 1: Test-set error on the lesion studies of Addemup. Due to the inherent similarity of each algorithm and the lengthy run-times limiting the number of runs to a ten-fold cross-validation, the di erence between the lesions of Addemup is not signi cant at the 95% con dence level.... In PAGE 92: ...show that more diversity is needed when generating an e ective ensemble. There are two main reasons why I think the results of Addemup in Table1 0b are especially encouraging: (a) by comparing Addemup with Regent-combined, I explicitly test the quality of my tness function and demonstrate its e ectiveness, and (b) Addemup is able to e ectively utilize background knowledge to decrease the error of the individual networks in its ensemble, while still being able to create enough diversity among them so as to improve the overall quality of the ensemble. 5.... In PAGE 119: ... Table1 2: Domain theory for the arti cial chess problem. The OR in the following rules means that only one of the two antecedents needs to be satis ed in order for the rule to be satis ed.... In PAGE 124: ... Table1 3: Domain theory for nding promoters.... In PAGE 128: ... Table1 5: Domain theory for nding ribosome-binding sites.... In PAGE 130: ... Table1 6: Part 1 of the domain theory for nding transcription-termination sites; see Table 17 for the remainder of this theory.... In PAGE 131: ... Table1 7: Part 2 of the domain theory for nding transcription-termination sites; Table 16 contains the rst half of this theory. % A region upstream from the site should be rich with A apos;s and T apos;s.... In PAGE 131: ... loop-even :- @-2= quot;TTTT quot;. % As in Table1 6, this rule encodes the notion that a longer stem is more stable. Like % before, the consequent of this rule, stem-even, is true if the sum of the \weighted quot; % true antecedents is greater than 6.... ..."

"... In PAGE 9: ... [104] also does an excellent job of distinguishing \traditional AI quot; from the study of autonomous agents (what we apos;ll term agent AI). Table1 summarizes the main di erences traditional AI and agent AI. [104] discusses the basic problems AI researchers have with adaptive autonomous agents, which includes interface agents.... ..."

"... In PAGE 10: ... The advantage of AI is that it considers the time factor. Table12 is the AI matrix, which is calculated by AD matrix (see Table 9) and page staying time (see Table 8). In this example, we use two-way discretization method, setting k = 2.... ..."

"... In PAGE 11: ... Here we explain how the AIS works by putting a node equipped with the AIS in the network scenarios typical for invocation of certain blocks or concepts. For the details, please see the pseudo code of the six AIS building blocks (Appendix B) and the values of the used parameters ( Table1 in Section 5). Bootstrap.... In PAGE 18: ... Clustering is used in all the experiments, as we already have shown its advantage over simple matching in [2]. The values of the system parameters used in the simulation are given in Table1 . The same default values are used in all the experiments.... ..."

"... In PAGE 6: ...As illustrated by ICU monitoring and summarized in Table1 , AIS niches are considerably more demanding than the niches occupied by typical AI agents. First, AIS niches require performance of several diverse tasks, sometimes concurrently and often interacting.... ..."

"... In PAGE 5: ... It then applies both the transfer advice and the user advice to learning in the target task. Table2 summarizes the AI2 algorithm in high-level pseudocode. Figure 2 illustrates the transfer part of this algorithm with an example from RoboCup.... ..."