Testing Agent Behavior
239
8.2 Statistical Testing of Data
Statistical testing procedures give invariants such as tools like DAIKON
[58], which help identify the upper and lower bonds of the resulting agent
variables through simulations. Adra et al. [1] used multi-objective optimization
to generate sample starting conditions for models to simulate. The iterations
were then sifted through, using DAIKON, to find the patterns of each memory
variable. The rules produced were able to give maximum and minimum values
for each variable, helping to find if any memory went out of allowable bounds.
This showed that there were some discrepancies in the model code.
Results can be checked by studying the outputs produced such as graphs
to help group large datasets into single pictures to find outlying data. Various
evolutionary testing methods can generate test sets to test the outputs. This
procedure involves using a computer program which uses genetic algorithms
to vary the test code, changing the bounds according to the data produced.
The resulting test case is the general test which can then be used to test the
system.
FIGURE 8.2: Screenshot of Weka analyzing cancer output.