Alright, I got my first ML algo (and ML project in general) up and running locally, which is pretty exciting. I trained it on ~2000 replays, although it’s clear that it’s confined to copying Demux-like algos due to the hard coded round 0/1 I provide it.
The algo is designed to just provide to me the defensive placements that it thinks would be best. I work the rest of the magic to place the best attack, so it’s certainly not operating “all on its own.”
Here’s the first game, which I think is pretty successful and shows it able to express ideas related to how to add on to Demux: second_game.replay (5.8 MB)
Here’s another, in which I don’t give it the first two rounds to give it structure. It’s a hot mess and seems to want to copy some dynamic algos briefly before it spirals out of control: third_game_organic_placements.replay (2.1 MB)
Overall I think the algo works pretty well. It’s certain a bit “unfair” to pure ML algos to hand mine optimized attacks and a well defined structure, but maybe later on I’ll get it to operate a bit more autonomously. I’ll see what I can do about getting this running live, but I suspect the dependency workaround will not install in time to be functional. I might need to wait it out for Python dependencies to be supported, unless I figure out how to make my own forward propagation. Update: “Unable to find ‘keras’” when I use the workaround…
From here I might just commit it to learning from Demux-algos. Ideally I’d like to set up some kind of self-reinforcement environment, but with the struggle of setting this up that might take a while…