MountainCar benchmark using TreeController

Starting the experiment:

./train

This will perform a learning experiment on the MountainCar plant using the 
Fitted Q Iteration algorithm. The Q function is approximated by 'Extremely
Randomized Trees' (ExtraTrees), one for each discrete action. After every 
episode a shell script (test) is executed, which evaluates the current 
contoller on a fixed set of starting positions. Results are written to a 
text file (performance.stat). 
