Well, I made the changes I identified yesterday, and got everything running, but I’m hitting issues training. I’m not sure if it’s not enough cycles, the structure of the network, or something else, but this feels like the point at which you either learn from others or learn from long and painful experience. It’s also pretty slow to run, given that it’s single threaded and in Python, so not exactly the best for iteration when each cycle takes minutes even with a trivially small network.

I do need to think through how to make traversal more efficient though. It balloons out across the network way more than I want it to. Again, this could be network structure; maybe there are just too many cycles created from my generation logic. Maybe I’ll look into this tomorrow, unless I decide to dive into research heavily. Also need more tests, but I kinda wanna keep riding the flow and only go back to that when I get stuck or need a little mental break. Onward…