Hier eine Beschreibung auf Discord:
„My A/B helper branch (based on killerduckys leelafish) is usable now. Since Wasp was the engine that Leela had trouble with in early days in TCEC, I thought it would be nice to try cooperating instead of fighting with Wasp, so I hooked up Wasp 4.5 as the A/B helper, did some trial and error testing to find a good balance between quality, speed and strength of influence on Leelas search and got something that looks far better than Leela on her own (T60 ~ 8KNPM on an RTX 2060).
`Score of lc0 master vs lc0 + Wasp 4.5 (4 threads, depth 12): 10 - 22 - 55 [0.431] 87`
Having that said, I'm now on to try improving the integration of the A/B helper into Leelas search. Killerduckys code works by selecting nodes in Leelas tree which already have got more visits than a threshold, these nodes are evaluated by the A/B helper (or as many of those nodes as the A/B helper has time enough to evaluate to a predefined depth). The PV of A/B helper is then used like this: the starting moves in the PV get their policy boosted in Leelas search. The drawbacks of the method are
- moves that never get enough visits by Leela will never be evaluated by the A/B helper
- only moves that are alredy in Leelas tree can get their policy boosted.
- policy boosting does not guarantee any level of exploration.
I'm keen to develop a completely different way of influencing leelas search: directly adding nodes into her tree. An A/B helper can provide a depth 15 PV in a few 1/10 seconds (on my old and slow system), ideally we would simply queue all those nodes for NN-evaluation in one go, and add them to the tree before backpropagating (adding them require handling them in a different order, deepest last, than backpropagating, which require deepest first). For slow but high quality nets, like 40b, I think a speedup in search could make a huge difference. https://github.com/hans-ekbrand/lc0/tree/a/b-helper