TalkChess.com

By the way, in your example, I think there also exists many equivalently short lines to get to that position without involving double blunders, so it may not be the case that the only chance of expanding the node is by one particular bad sequence of moves that never worth exploring. Sure, this was ...

Or do you mean we still expand it when the main line wants, but pretend the move history was from the shorter line so to speak? It is not a depth-first search, you do not need to always expand main line nodes to make progress, the search progresses by increasing overall expansion rate and that impr...

That's not what I meant, I questioned the breadth firstness here. What if I get to the position with a long line first and expand it a bunch, then hit it with a shorter line? Also apply rule #2 and for the longer lines, only when your shorter line hits the node for you to expand it. This doesn't wo...

How is that contradicting? First they make a mistake then you make a mistake back to get to the position. Since your move was a mistake it will hardly be expanded since the better alternative for you would actually be winning. If this ever occurred then apply rule #2 and move on? I still don't unde...

But what if the shortest path doesn't want to expand it because the shortest path included an opposition mistake which gives us a much better option to focus on? But in the main line where this position shows up later this is the main move which we want to expand a lot? Well that is contradicting: ...

I don't quite understand... Very simple: 1. Only the path of globally shortest occurrence of a node gets the right to explore its child. 2. On other occasions treat it as a terminal node and adjust the score locally if the path indicates a repetition. But what if the shortest path doesn't want to ex...

A few people are working on an improved version of Leela's search right now which allows better exploration and among other things transposition handling. In practice that would just be a search tree in memory as usual, however different positions' child node pointers might point to the same node if...

That's an interesting position. My freshly implemented perft worked for all positions on the cpw except this one where on perft 4 it gave exactly 4 nodes too many which seemed like an oddly small number to be off. As it turns out I somehow didn't clear e.p. rights when castling which gave the exact ...

Not sure if this has been suggested like that before, but I just had an idea for book creation without a large set of games behind it already. Thinking about it further it is actually somewhat similar to Leela's search so I might as well describe it as such. So in the book creation step you basicall...

A friend of mine was asking whether there have been serious attempts to find good handicaps for various rating differences. In engine games there seems to be some rough understanding of it, however what about human vs human games? Maybe @lkaufman has some insights? Yes, I know a lot about that topi...