I continued training my main model, and the results seem pretty nice. In addition I've updated the method I use for exporting the jit Pytorch models I'm using in the game. This does seem to bring them closer to a 1:1 adaptation of the original models. Still not perfect but its possible that when I start training models with larger training datasets it may start to close that gap. I will try and make a Windows update to match the Mac version soon.

I've been experimenting with a smaller 3 layer model lately as well. I don't have it included in this version, though. I'm hoping having smaller models mixed in with the larger models will allow having more AI opponents on the map at once. That currently isn't really an issue, I can run ~9 opponents on my system now (~15 if I mix in the smaller model) and that makes the map decently lively. As I scale the map size up though, having more enemies in the game will be crucial to keep the game from feeling empty.

In addition, in the 1.0.2 / 1.0.3 updates I’ve added the effect of the ships breaking apart upon destruction. It is pretty nifty. I need to work on the explosion effects, spice them up a bit.

--- In the works, I have an update that includes firing bullets as a weapon addition, and 'prizes' to collect for rewards, and shielding you can build up from the prizes. The bullet settings are fixed for now, but I plan to make all the weapons systems more variable with the addition of ship upgrades. Prizes spawn randomly on the map, and also drop when a ship is destroyed. I've began training a model with the new updates, and it seems to be going well in my testing of it. It is a little early to tell, and I think I'm going to need ramp up the amount of training data I'm using, but they fly around and shoot at least. Training can take weeks for me, but I'm happy with it so far.