Open hwj-111 opened 2 years ago
Updated plot on 2022/02/10
I found now cgos BayesElo has accumulates enough b40v50 data points. So I plot both v100 and v50 together. v50 fitted line (slope ~ 0.15) seems lower by 180 elo than the v100 fitted line (slope ~ 0.19). This value seems a bit larger than the benefit of purely doubling the visits (+125 elo), indicating v100 is much preferred than v50 when using 40b networks (I guess, more visits needed before reach the 40b saturated limit)
Here is the plot of v100 (red) and v50 (green) of 60b network progress. The delta between them is about 170elo. The slope is much smaller than 40b ( only about ~0.14 per network# for v100, and 0.05 for v50!), indicating the progress is slower (when evaluate progress on v100). Maybe, v200 or v400 is much better ( large slope)?
I have ran some katago 60b nets on CGOS with visits/move = 200 (blue dots). Their progress is clearly better than v=100 (red) and 50 (green). Maybe, to see the actual progress, visits/move has to be higher, at least, 200 or so
update for katago 60b progress on CGOS, with vistts/move = 200, 100, and 50
I plot the progress trend of 40b network based on the data on cgos beyes elo Roughly improved at +0.2 elo / network