search

kmcrage / leela_lite

A python implementation of lc0 by dkappe, ideal for testing new algorithms.
GNU General Public License v3.0
1 stars 0 forks source link

mcts-ab #3

Open kmcrage opened 5 years ago

kmcrage commented 5 years ago

http://talkchess.com/forum3/viewtopic.php?t=66886 http://se1f330a320707f8e.jimcontent.com/download/version/1467247204/module/12396903227/name/a%20rollout-based%20search%20algorithm%20unifying%20mcts%20and%20alpha-beta.pdf

kmcrage commented 5 years ago

huang ab search done: with k=5 its a bit weaker than uct

kmcrage commented 5 years ago

In uct, we could use Q = (total_value/num_visits + v_plus(depth-1))/2 but then do we halve cpuct too? In a combined search, we want to minmax over at least 5 children, but all with visits>0.

kmcrage commented 5 years ago

or maybe Q=min(v_plus, total_value/num_visits) ie the lower of the upper bounds (or maybe the max) In the opening, we want uct to dominate, but in the endgame we want ab to dominate.

or a clamp between v_minus and v_plus