\addvspace {10\p@ }
\addvspace {10\p@ }
\contentsline {figure}{\numberline {2.1}{\ignorespaces Features of the COCOMO model ontology.}}{10}
\contentsline {figure}{\numberline {2.2}{\ignorespaces COCOMO 1 effort multipliers, and the sorted coefficients found by linear regression from twenty 66\% sub-samples (selected at random) from the NASA93 PROMISE data set; from\nobreakspace {}\cite {me05a}. }}{13}
\addvspace {10\p@ }
\contentsline {figure}{\numberline {3.1}{\ignorespaces Four steps of case-based reasoning, from \url {http://www.peerscience.com/intro_cbr.htm}.}}{15}
\contentsline {figure}{\numberline {3.2}{\ignorespaces The Brooks' Law Query for the NASA93 dataset in COCOMO II format.}}{19}
\contentsline {figure}{\numberline {3.3}{\ignorespaces RELEVANT= cases nearest to $context_1$.}}{20}
\contentsline {figure}{\numberline {3.4}{\ignorespaces $Best$ (top) \& $rest$ (bottom).}}{21}
\contentsline {figure}{\numberline {3.5}{\ignorespaces Rank with Equation\nobreakspace {}3.1\hbox {}.}}{22}
\contentsline {figure}{\numberline {3.6}{\ignorespaces A $K_1=20$ neighborhood of $context_1$ (NASA93ii train set).}}{23}
\contentsline {figure}{\numberline {3.7}{\ignorespaces All rows of Figure\nobreakspace {}3.6\hbox {} satisfying $R_1 : pmat=3$. }}{24}
\contentsline {figure}{\numberline {3.8}{\ignorespaces The testing set with all cases not containing $pmat = 3$ removed. }}{24}
\contentsline {figure}{\numberline {3.9}{\ignorespaces Result of applying the learned constraint $pmat = 3$ to the Brooks' Law query $q$ during testing. The median estimate reduction from 235 to 81 represents a 66\% reduction is software effort by applying $pmat = 3$.}}{24}
\contentsline {figure}{\numberline {3.10}{\ignorespaces Revising $q$ to learn $q'$.}}{25}
\contentsline {figure}{\numberline {3.11}{\ignorespaces {$\mathcal {W}$}2's syntax for describing the input query $q$. Here, all the values run 1 to 6. $4\le cplx \le 6$ denotes projects with above average complexity. Question marks denote what can be controlled- in this case, $rely,time$ (required reliability and development time)}}{26}
\addvspace {10\p@ }
\contentsline {figure}{\numberline {4.1}{\ignorespaces Seven data sets from \url {promisedata.org/?cat=14}: {\em effort} is total staff person-months; {\em time} is calendar time (start to stop); {\em defects} represents the number of delivered defects. }}{28}
\contentsline {figure}{\numberline {4.2}{\ignorespaces Example project $controllable$ file for Chinese software projects after discretization. Ranges were assigned randomly for this project. A ``?'' represents a controllable feature. If an attribute range isn't specified in the project, it is ignored.}}{30}
\contentsline {figure}{\numberline {4.3}{\ignorespaces Average execution times for the W and W2 algorithms. By removing the $O(n^2)$ kth nearest neighbor calculation from W we drastically improve performance, especially on larger datasets such as China (499 cases).}}{31}
\contentsline {figure}{\numberline {4.4}{\ignorespaces Performance of W2's Overlap relevancy filtering vs W's kth nearest-neighbor filtering for 5 unique datasets.}}{31}
\contentsline {figure}{\numberline {4.5}{\ignorespaces Effort estimation improvements ($100*\frac {initial-final}{intial}$) for five unique datasets. Sorted by median improvement. Gray cells represent no improvement in effort estimates.}}{33}
\contentsline {figure}{\numberline {4.6}{\ignorespaces Effort results for five non-COCOMO datasets. }}{34}
\contentsline {figure}{\numberline {4.7}{\ignorespaces Range of changes in median and spread generated by applying the recommendations of {$\mathcal {W}$}2. The median observed changes were (20.5, 20.5)\% for (medians, spreads), respectively. }}{35}
\contentsline {figure}{\numberline {4.8}{\ignorespaces Recommendation frequency across 20 runs of {$\mathcal {W}$}2 for reducing individual goals ($defects$, $effort$, or $months$) as well as all goals at once ($all$).}}{36}
\contentsline {figure}{\numberline {4.9}{\ignorespaces Examples of drastic changes to software projects.}}{38}
\contentsline {figure}{\numberline {4.10}{\ignorespaces Comparing defect, effort, and month estimation reduction percentages ($100*\frac {initial-final}{intial}$ of drastic business decisions vs {$\mathcal {W}$}'s recommendations for the Ground case study.}}{38}
\contentsline {figure}{\numberline {4.11}{\ignorespaces Comparing defect, effort, and month estimation reduction percentages ($100*\frac {initial-final}{intial}$ of drastic business decisions vs {$\mathcal {W}$}'s recommendations for the Flight case study.}}{39}
\contentsline {figure}{\numberline {4.12}{\ignorespaces Comparing defect, effort, and month estimation reduction percentages ($100*\frac {initial-final}{intial}$ of drastic business decisions vs {$\mathcal {W}$}'s recommendations for the OSP case study.}}{40}
\contentsline {figure}{\numberline {4.13}{\ignorespaces Comparing defect, effort, and month estimation reduction percentages ($100*\frac {initial-final}{intial}$ of drastic business decisions vs {$\mathcal {W}$}'s recommendations for the OSP2 case study.}}{41}
\addvspace {10\p@ }
\contentsline {figure}{\numberline {5.1}{\ignorespaces Contexts of 4 case studies. \textit {\{1, 2, 3, 4, 5, 6\}} map to \textit {\{very low, low, nominal, high, very high, extra high\}}.}}{43}
\contentsline {figure}{\numberline {5.2}{\ignorespaces Pseudocode for SEESAW}}{48}
\contentsline {figure}{\numberline {5.3}{\ignorespaces Example of SA's forward and back select.}}{50}
\contentsline {figure}{\numberline {5.4}{\ignorespaces Number of times algorithms were top-ranked (largest is 4: i.e. one for each Figure\nobreakspace {}5.1\hbox {} case study). }}{52}
\contentsline {figure}{\numberline {5.5}{\ignorespaces Changes in median and spread for the NASA93 dataset. }}{54}
\contentsline {figure}{\numberline {5.6}{\ignorespaces Changes in median and spread for the COC81 dataset. }}{55}
\addvspace {10\p@ }
\contentsline {figure}{\numberline {6.1}{\ignorespaces Relative effects on development effort. From\nobreakspace {}\cite {boehm00a}. }}{59}
\addvspace {10\p@ }
\addvspace {10\p@ }
\addvspace {10\p@ }