Learning Track Stuff

[Erek Speed]

Hi,

I've come out of the shadows to update my agent and submit but just
wanted to bring up a couple issues for people running the simulator
for learning.

1) Right now the visualization is set to on for training which is
probably not what you want so pass the proper argument or set it
manually to avoid for anything but the final run.

2) numberOfTrials is set to 1000 in LearningTrack.java which is fine
but this isn't used. LearningTask.EVALUATION_QUOTA is what is used
and that is 100000 which is big.

3) The fitness function is strange. I'm using the Level 1 arguments
and getting 38000 for a run which dies half way through. Later I
trained a guy up to 79000 and it was a quick death. It must be
something accumulating across multiple runs but I can't see what it
would be right now. I'll look at it again tomorrow.

Cheers,
Erek

[Slawomir Bojarski]

I believe I've tracked down the cause of #3. Whenever the environment is reset before a run, the "resetStatic" method is called on the Mario class. This method resets some static variables that are factored into the computed weighted fitness. The problem is that one of these variables ("greenMushroomsDevoured") is not being reset. The value of this variable accumulates over the runs and creates an ever-increasing score for the agent. This seems to only happen in "level 1" in the LearningTrack class, but affects scores of latter levels if you're running through multiple levels.