lkaufman wrote: ↑Sun May 12, 2019 8:23 pm
konsolas wrote: ↑Sun May 12, 2019 6:23 pm
lkaufman wrote: ↑Sat May 11, 2019 11:48 pm
~snip~
The scores reported by an MCTS engine like Komodo MCTS already reflect the possiblility of the opponent going wrong in plausible ways. Maybe it doesn't do it the same way you would, but at least to a significant degree it does what you want.
That's really interesting. If this is possible with MCTS, have you considered adding a mode to Komodo which adds more weight to the possibility of the opponent going wrong, which would allow people to find trappy lines in analysis?
If regular Komodo did that, it would just become a much weaker cousin of Komodo MCTS. That's the fundamental difference between standard ("Alpha-Beta") engines and MCTS engines; standard ones assume that the opponent will always play the move the engine considers best, whereas MCTS assumes that all reasonable moves have some chance of being chosen. Komodo MCTS (and Lc0 and spinoffs if you have suitable GPU) are the engines you should use.
I think that’s a bit cart before horse. Both MCTS and AB “assume” all (reasonable) moves have some chance of being played, else they would not be searching them.
AB creates and examines a tree like a little mouse that runs up and down branches examining the leaves on a sort of start at the left and end at the right basis. AB finds and outputs a PV which says: here’s a branch down the tree which picks out an optimal path to a tree leaf with a score that my opponent can’t stop me from achieving.
MCTS make single foray stabs into the tree. It stabs down a line, gets to the end, evaluates, and returns to the root. Then another stab, and so on. Stabs go into regions of the tree. MCTS finds an outputs a move which says: if I play this move, my opponent has several reasonable replies, and I have some reasonable replies to his reasonable replies, and so on, where I found that my stabbing forays, on average, in this particular region of the tree, led to leaves where I was winning more often than my opponent.
Richard Lang, a long time ago, likened search (which was AB) to picking a path through a minefield in order to reach the pot of gold at the other side without stepping on any mines. That’s a good summation of AB search.
By extension, MCTS search is about entering a minefield with disparate regions, where some mines are set to blow you up, and some set to blow up your opponent. MCTS avoids the regions full of mines that blow it up, and steers to the regions where are the mines that blow up its opponent.
So, AB is on the lookout for tactical pathways and MCTS plays positionally until its opponent position just disintegrates all by itself.