If math wouldn't be my profession, I probably wouldn't dare to speak so freely. Though I have to admit, that I did some programming since then.
While you might be right about that, it sounds complicated and seems less useful than a basic answer on plain corner attacks.unhandyandy wrote: ↑Wed Feb 06, 2019 8:15 pmYes, you could use retrograde analysis to develop an edge-fort engine.Are you talking about edge forts? That might totally be. I was using that technique for edge forts during my games.
The real truth about this is that I don't know how to go about it. Furthermore it's hard to test, because the edge fights are just a result of massive weaknesses black created during defending corners. So until we understand how these corner attacks work, it's hard to determine which kind of edge forts are interesting.
At the same time you are right about the fact that a naive "just do some machine learning without prior knowledge", is way to naive to get a result for edge forts in a reasonable amount of time. It's just to hard to construct one "by accident". I was planning to leave that issue for later, but if you have any ideas I'm all ears.
We always have graph based based pruning techniques.unhandyandy wrote: ↑Wed Feb 06, 2019 8:15 pmDidn't the original AlphaGo program also do local tactical analysis separately from global strategic analysis?
Afair there were never a special part for tactical or strategical results. At the same time we have two different outputs (that were attached to entirely different networks in the first version), while one looks for "interesting moves" (in particular all tactical ones) and the other one is trying to give an estimation who has the upper hand (essentially giving a strategic judgement).