Changes between Version 4 and Version 5 of AdditionalMaterial/ECMLPKDD
 Timestamp:
 06/28/12 12:43:52 (12 years ago)
Legend:
 Unmodified
 Added
 Removed
 Modified

AdditionalMaterial/ECMLPKDD
v4 v5 19 19 == Identification of NonLinear Models == 20 20 As a preparation for symbolic regression a number of parameters have to be configured. A very important parameter is the grammar that restricts the possible shapes of evolved models and defines the basic building blocks for symbolic regression models (function set). 21 [[Image(grammar.png )]]21 [[Image(grammar.png, with=300)]] 22 22 23 23 Other important parameters are the error function that should be minimized (e.g. mean of squared errors, mean of absolute errors, Pearson's R², ...), and the maximal size of the models. Additionally the parameters of the underlying evolutionary algorithm like population size, mutation rate, and number of iterations can be configured. … … 27 27 28 28 In HeuristicLab a number of different charts and error metrics are available directly in the GUI for each produced solutions. All results are updated dynamically while the algorithm is running. 29 [[Image(screenshots.png )]]29 [[Image(screenshots.png, with=300)]] 30 30 31 31 == Identification of Relevant Variables == 32 32 Frequently it is not necessary to learn a full model of the functional relationship but instead only find a set of relevant variables for the process. This can be achieved easily with HeuristicLab through analysis of relative variable frequencies in the population of models. The following Figure shows a variable frequency chart that clearly shows the six most relevant variables. Notably, the relevance of variables is determined based on nonlinear models. So, nonlinear influence factors and pairwise interacting factors can be identified as well. 33 33 34 [[Image(relevantvariables_small.png )]]34 [[Image(relevantvariables_small.png, with=300)]] 35 35 36 36 == Simplification of Models with Visual Hints ==