I'm having an issue with the bot timing out during initialization. I'm not super familiar with the Python run-time. Is there some parameter i can modify locally in the game to increase the allowed init time or is this a symptom that I've missed something in the set-up process.
Use the -t flag to disable timeouts.
just wondering how long it takes to train the bot. Mine is taking a really long time.....
What do you call " a really long time" ? Are you running the computations on your gpu or cpu?
For reference, I'm training on a GPU and it takes about 35 hours to train. I'm not using the starter code though, so the setup might be different. I feel like mine should be faster than it is though... still working on making convergence faster.
Training my bot takes about 2h on gtx 1080. It converges a lot earlier when using batch normalization.
Wouldn't it be preferable, rather than introducing further parameters, to do self-play for the RL phase (after perhaps initializing Q with supervised learning against nmalaguti)? That's essentially the approach taken by AlphaGo. You'd be likely to get a distribution of wins and losses and could apply the TrueSkill reward/penalty to more directly target how the bot will be scored.
@brianvanleeuwen are you able to define a set of rules as well for moves or override. For example, don't move if square.strength < target.strength, even if the ML says so.
Side Note: I am unable to run the whole replay.zip and keep getting memory errors. I can run ~800 replays, on 8gb of ram with 64bit python, is that normal?
I originally tried running this model on an m4large ec2 instance with 8 GB RAM, and it also errored out on the full 2096 replay set. I started it up on an m4xlarge instead, and it's consuming about 9.5 GB of memory.