The learning uses ReinforcementFeedback? to work out the optimal solution, and it tries to balance Greed? with Exploration?. Currently I aim only for TestBot to learn how to move in 1v1 so that it survives the most. It doesnt do too badly currently, but takes around 100 rounds to get going.
Its targeting is pretty bad currently (shoot when I scan), so I might just drop in Sandbox mini's targeting so I can concentrate on the learning movment.
It does improve against top bots like Sandbox but unfortunatly not to the point where it is near beating them. :)
Check back for updates, and I might post some results as well. Not that anyone is going to read this! :)
Ok, TestBot is still VERY experimental - dont expect good results yet! :)
I think im going to call this bot AgentSmith. :)
Update: Request from Pez to test TestBot against Marshmallow:
As you can see, TestBot's learning isnt great yet, and doesnt appear to work at all against Marshmallow! :)
Ok, tests against Sandbox
As you can see it gets slightly better against DT, but reaches a peak of managing to take about 1/3 of the battles from DT.