Training runs

Overview

Run #	Reference	Summary	Currently Active	Net Numbers	Best nets
NA	Old Main	Original 192x15 "main" run	No	1 to 601	ID595
test10	Lc0 Transition	Original 256x20 test run	No	10'000 to 11'262	11250 11248
test20	Training run reset	Many changes, see blog.	Yes
test30	TB rescoring	Experiment with network initialization strategy, trying to solve spike issues. Experiment with Tablebase rescoring	Yes

Technical Info and sampling ratio

Most data from this sheet

• Alpha Zero reference paper Use best guess for games length and assuming resign cuts game length by 30%

Item	Without Resign	With resign
Positions per game	135	95
Total games	44,000,000	44,000,000
Total positions generated	5,940,000,000	4,158,000,000
Batch size	4,096	4,096
Total steps	700,000	700,000
Total position trained	2,867,200,000	2,867,200,000
Sampling ratio	0.48	0.69

• Old Main Initially new networks generated based on fixed timing rather than on games

Item	up to ID xxx	From ID xxx	From ID yyyy
Positions per training game		135	95
New networks per day
Training Games per day
Training Games per network
Total training games
Total positions generated
batch size
total steps
total position trained
Sampling ratio	****	0

Original main pipeline				

games per day 160,000 positions per day 21,600,000

networks trained per day 6 batch size 1,024 Note: batch size changed from 1024 to 256, but we're not sure exactly when. steps per training session 50,000 positions per training session 51,200,000 positions used for training per day 307,200,000

sampling ratio 14.22

New main pipeline	Test pipeline		Test 2018-07-20 ? Not sure about what the current values are	

games per new network 40,000 40,000 70,000 positions per game 135 135 82 positions per new network 5,400,000 5,400,000 5,740,000

batch size 256 2,048 2,048 steps per training session 10,000 2,500 2,500 positions per training session 2,560,000 5,120,000 5,120,000

sampling ratio 0.47 0.95 0.89