Increasingly modeling is being used in all the
software life cycle stages studied in the field
of software engineering.
By using models from two such stages, 
the requirements stage and the testing stage,
we show that our machine learning based 
metaheuristic search technique
can be used to optimize these models.
This technique is called Iterative Treatment Learning, ITL.
We apply ITL to three NASA cost/benefit models 
used during the requirements stage and
two biomathemitical and three NASA flight models
used during the testing stage.
These last five models are investigated
using an integrated development and testing framework,
SPY, that has built-in ITL search capabilities.

These studies improve upon previous work done
with ITL by Menzies \etal by
1) discussing ITL's characteristics in metaheuristic search terminology
2) increasing the number and complexity of the models studied
3) exploring the option space of ITL, including a new discretizer and
4) investigating the stability and variance of ITL's solutions.

With the cost/benefit models we show how our new
method for discretizing the model's dependent variables
1) outperforms the previous discretizing method,
both on solution quality and convergence speed
and 2) outperforms a simulated annealer, a commonly
used metaheuristic search technique, on 
convergence speed while finding the same quality
solutions.
With the biomathematical models we show that SPY
can find potentially useful range restrictions
to model inputs that restrict the model's output 
to specified behavior modes.
Finally, with the NASA flight models we show 
that SPY can verify temporal properties
in models with real valued inputs, comparing
SPY's performance to a commercially available tool,
Reactis, that uses a random and heuristic search
strategy.