Firstly, I uploaded 50 moves of my bot playing itself on a 19^2 board over on L19.
Because I am not building this for high performance computing, playouts are going to be fairly sparse.
Because playouts are sparse, heuristics for choosing what next moves to compare are extremely important. Initially I was hoping to avoid writing in my own heuristics and introducing bias, but I need to examine alternatives to sheer volume of playouts.
Sofaras heuristics, I want to convolve (Wikipedia) conventional 'good' (and bad) go shape kernels to find good shape candidates for next moves. But the implementation is not exceedingly clear. I am not totally sure that convolution is the fastest way of finding these. A friend tells me four channel convolution is well optimized (whereas I would normally just break a problem into 2d convolution). Semantics. Will report back.
Aside from this, I want to utilize graphics processing (convolution on graphics processors is normal).