2022-02-13
Training of 20x256 networks is on-going!
Currently for each generation there are at the least five separate training runs from last promoted network. Each run is done with slightly different meta-parameters (e.g. number of generations, learning rate, etc...) and produces two candidates. The second candidate of each run is actually the weights' average from the last few training steps and is indicated as -swa in the name tooltip.
Latest SAI release: 0.18.2
The current version of the SAI engine is still the one released an year ago. We are currently working on an slightly improved version that should come out soon.
Among the current engine features there is:
New fpu estimate for unvisited nodes.
Improved performance for agents with non-zero lambda/mu
Better resigning and fixed score/komi in handicap games
More verbose stats after think
发表于 2022-2-23 13:56
收藏
分享
顶
踩
提升卡
喧嚣卡
变色卡
发表于 2022-2-23 22:51
