Open harbecke opened 4 years ago
is quickfixed with fd76242 - no more elo calculations
solving the problem seems to be difficult, i took a quick look at https://cran.r-project.org/web/packages/BradleyTerryScalable/vignettes/BradleyTerryScalable.html which is based on https://arxiv.org/abs/1011.1761 but that doesn't solve that each run takes a long time
maybe we should run the elo calculation once at the end of the training to determine the best model? @cleeff @simonant any ideas?
Hmm that's unfortunate. I remember that I encountered this already and in elobayes[1] also an incomplete game graph is allowed. You can have each agent play a fixed number of games against a random subset or the best agents so far. The numbers aren't very accurate in that case but probably good enough.
Having agents being evaluated 'at the end' / after training sounds like this could easily be missed if you want to intermediately abort training.
[1] elobayes is the tool we previously used and which is also used in the AlphaGo papers.
sorry for being unclear. we already have an incomplete graph (we let each new model play against the best m=3 models so far), however the calculation of the elo list itself takes a very long time (we have up to 5000 models during a complete training run).
i think this is a low priority issue, as the reference models are more informative for the progression anyway.
Ahh that makes more sense. Yes, I totally forgot about that. I agree let's keep the elo computation out for now.
hex/elo/elo/create_ratings takes approximately
n²/10000
seconds to calculate forn
models on the server. this is way too much for long trainings and should rather be an approximation than quadratic time.