|
|
Discovery Node |
The wizard-driven discovery node is DATALAMP's key feature, offering rule discovery using a simulated annealing algorithm. The discovery node runs in one of two modes: single experiment and multiple (batch) experiments. In both modes you set problem-specific parameters, including the limit on the number of conditions in the generated rule, the target class and the required accuracy threshold.
Although sensible and generic defaults have been set on the SA engine, a large number of advanced options are available for expert users. The advanced options include setting the initial temperature, terminating a temperature stage, selecting the cooling schedule, operator selection methods (such as random, probabilistic or adaptive) and engine termination.
At the end of the experiment, which you can terminate early at any point, the best rule found is copied to a scratch pad (similar to the Windows clipboard); you can paste the rule back into a rule node at a later point.
In multiple experiment mode, you can set a range for each of the problem-specific parameters. Each set of values becomes an experiment and the engine performs each experiment in turn; this option is useful for determining sensible values of lambda and is generally left running overnight.
Full details of the options available for the discovery node can be found starting with the general options page of the discovery wizard.
During execution of the discovery node, a dialog shows the progress of the search:
The graph on the dialog shows the last 100 accepted solutions in green and the best solution in red. If the dialog does not update frequently, or the "Not enough data" message is shown, see the troubleshooting section on discovery node progress graph does not frequently update.
