By Lothar Jung
Date 2022-03-07 12:28
Edited 2022-03-07 12:33
Upvotes 1
Ich würde sagen, er modifiziert Lc0 29 dev, um die Suchergebnisse der A/B- Helper Engine in die CPUCT Suche von Lc0 einfließen zu lassen.
Siehe GitHub Einführung:
„When Vanilla Leela decides which edges in the search tree to extend, she traverses the tree from root and applies the CPUCT formula to decide which branch to go down, until a leaf is found and the edge is then extended, i.e. the node that represents the position that the move represented by the edge leads to is created, and unless that node is terminal or already in the cache the neural net is queried about that node. This version of Leela combines that way of choosing edges to extend with another way: edges representing moves recommended by another engine are also extended.
A small fraction of all nodes in Leelas search tree are sent to an auxillary helper (via UCI), and the returned principal variation(s) are used to choose edges to extend in Leelas search tree. For the system to be effective, the auxillary helper should be a CPU-based engine, and Leela should be running on a GPU.“