Hey, that's awesome you've been trying to use some gradient-based learning for Raven! I wrote a bit more about how I used it in BeepBoop/Understanding BeepBoop. Is it similar to how you are doing things? It's tricky to apply deep learning methods to robocode due to compute/Java limitations, but it's definitely an interesting thing to explore!
- Hi, it's in fact not so similar to what has been done with BeepBoop. It also is much simpler (and dumber) since I kind of rushed it :)
- I'll actually add a small section named PyTorch to Raven's page explaining the current method I use to tune since it is rather hard to explain it in a single message and it would be easier to keep track of :)
- Raven 3.58 was more of an experiment to see if APS would increase against the bots I tuned using rather than achieving general success. I agree that it is indeed very interesting to explore.
- It also seems to me that Java might be getting SIMD support in the future(judging by the unstable Vector class in the latest version?) and quantization + SIMD could make a huge difference in execution speed of Neural Networks.