carykh/PrisonersDilemmaTournament

How the heck is this gonna be actually run

cadyn opened this issue · 4 comments

cadyn commented

I assume at least a few of us here are going the route of NN where some computationally expensive matrix multiplication might need to be done to evaluate the results. I think carykh said it would repeat the tournament like 1000 times to help eliminate randomness, and that's ontop of every one of our agents playing every other agent. The sum of all of this is that this is gonna take a lot of computational power, maybe too much to be even practical?

See #20 for if you should worry about your own strategy in terms of time, but otherwise it will probably take a while but not too long. Depending on the amount of final strategies, it will probably take a couple weeks as is, but that should be able to be brought down quite a bit if multiprocessing and non-random strategy caching are added. #17

I'm using neural networks, and mine executes and returns in around 5ms, even with all the debug features enabled.

isn't using another library other than NumPy and random is not allowed? or building our own NN from scratch is ok?

isn't using another library other than NumPy and random is not allowed? or building our own NN from scratch is ok?

As long as you'r not importing another library, and you aren't reading-writing any files, and you're not stalling on purpose, anything goes. I have made a simple nn library that only uses random(And time, but only for the debug logs, which isn't necessary. I actually have a repo of this but please don't look at it) a while back, so one ctrl+c ctrl+v and a few modifications later, I got it working.