@ez4u:

my guess would be that its a default minimal value, which seems rather low:
around 20-25 is the threshold where a bernouilli distribution begin to be approximated by a normal distribution so roughly:
(here we consider that a given move give a probaility p that a randaom playout after the move will give a won game for B, and we want to approximate p by playing a limited number of playout

the error is of order 0.5/sqrt(23) which is approx 0.1. This seems is a rather big error bar.
it might be interesting to check if this min number of playout can be tune up and if it would affect the result.
having a 1% error interval means around 2500 playout.... which is a lot (to gain a factor of 10 accuracy you need 100 more playout). but it needs to be compared to the in depth exploration of the may vars: you have roughtly 300 plays looking "not interesting" so exploring them all at 2500 playout would be 750.000 play out which is still far less than playout for the top move.

So maybe fuego will run at roughtly the same speed with this value tuned up ?
Not sure, though, that it would bring a better playing strenght.

Does esmeralddemon as the time and willingness to give it a a shot for 1 move ?

If we're talking about sufficient data points to improve play beyond chance, then I suspect 1 move won't tell us much

Thank you for sharing your experiment. I agree that intuitively, this focus on one move is excessive. There are also some theoretical arguments against it. Still, we do not have anything that works better at the moment. This is definitely worth studying.

One problem with doing this in the whole search tree is that because of the child selection formula, exploring bad moves by doing more exploration costs you performance because of the averaging process. I.e. assume the following situation:
you play Black and have two choices:
move a is very aggressive, but it works and it will win the game immediately by some complex tactics.
move b is solid, and will lead to a quiet, balanced situation, say a .50 score.

To simplify, let's also assume that the opponent's reply to a, White a2, is forced. So let's look at the position when it is Black's turn to play after Black a, White a2. Since the position is very tactical, there is probably only one way for Black to win from here. With low exploration and deep search, the program may discover this win, and most simulations use this move, so the weighted average looks good and move a is selected. But with higher exploration, the program will explore many more of the losing continuations in this subtree, and the average value of a will drop.

I think it is hard (but of course very interesting) to try to fix this problem in general. A simpler approach would be to force more exploration at the root node only. I agree it must be wrong to have say 100 million simulations and spend more than 99 million on the same move.
Here, the tradeoff is between exploring the top move deeper, and exploring alternative moves deeper. I believe that at some stage MoGo was using a different bandit formula for the root node, and they may still do that. But I do not know what they do, or if the details are published. I will meet Olivier next week so I will ask him.

As far as I know no one ever tried different root bandits in Fuego.

Martin