Incomplete metaheuristic search has been empirically shown
to help find near-optimal solutions for problems
that defeat complete search techniques,
(?either because the search is too large, or
there is difficulty in developing an ordering
heuristic or an admissible cost heuristic?).
This thesis uses a metaheuristic search
technique to in eight different models.
The technique used, Iterative Treatment Learning (ITL),
has been previously developed by Menzies et.\ al.\ ,
but that previous work never explored the
characteristics of ITL.
In addition, ITL requires a discretizer
for the dependent variables, but all
previous work has used the same discretizer,
diagonal striping.

We used three NASA cost/benefit models to investigate
a new discretizer called extreme sampling and
several different parameter settings to this new discretizer.
We also present SPY, an intergrated model
development and search environment that uses ITL.
SPY uses a slightly different treatment learner
that has a built-in discretizer.
With SPY we investigate restricting models
to specified behavioral modes, using
two hand-written biomathematical models and 
three NASA flight models translated from the modeling
language Simulink.

The experiments with our cost/benefit models
show which parameter settings to
extreme sampling perform the best on these models.
Further, we show that the solutions found by ITL
are stable and have a low variance.
After tuning ITL, we show that our new discretizer
outperforms our old discretizer, on both convergence
speed and solution quality.
After tuning, ITL is also shown to find solutions
of a similar quality to simulated annealing,
a commonly used metaheuristic technique,
but an order of magnitude faster.
The experiments with our biomathematical models
show SPY's ability to find potentially useful
range restrictions to model input variables.
Finally, the experiments with our NASA flight models
show SPY's ability to validate temporal properties
in models with real valued inputs, with the same
competence as a widely used commerical tool, Reactis.