How are you defining winrate? For instance when you say winrate is that wins/games or (wins +.5draws)/games?
By Elo rating do you mean Lichess rating? Are they the same thing?
Can you add the winrate for Stockfish against Stockfish, for each data set? I'd like to get a sense of what the objective winrate is for each data set.
I missed a time control differentiation. For example, is the bishop pair really a advantage in a blitz or bullet game?
Some answers:
- This is blitz and longer, bullet is discarded
- Winrate is just winrate, not winrate + draw*.5
- Elo rating == Lichess rating, I’m using it in the colloquial sense, even though technically Lichess doesn’t use Elo
- This is blitz and longer, bullet is discarded
- Winrate is just winrate, not winrate + draw*.5
- Elo rating == Lichess rating, I’m using it in the colloquial sense, even though technically Lichess doesn’t use Elo
@boxbeatsy said in #4:
> Can you add the winrate for Stockfish against Stockfish, for each data set? I'd like to get a sense of what the objective winrate is for each data set.
This would require a ton of compute, since there’s like 50 million games in the dataset. Might do a different analysis sometime later that includes Stockfish evals, but it won’t be this data set for practical reasons
> Can you add the winrate for Stockfish against Stockfish, for each data set? I'd like to get a sense of what the objective winrate is for each data set.
This would require a ton of compute, since there’s like 50 million games in the dataset. Might do a different analysis sometime later that includes Stockfish evals, but it won’t be this data set for practical reasons
@marcusbuffett Ty for the reply and interesting content! On the topic of sample size, I believe just running the games out on a random sample of 1k games per data set would result in sufficiently narrow confidence intervals. For example, if the observed winrate is 60% over 1k games, the standard error is only ~1.5% ( (60% * 40% / 1000) ^ 0.5 )
Interesting idea to measure this :)
This is a very good blog! I very rarily give likes, but I liked your study a lot! Most of it I already knew, like that beginners struggle to convert two minor pieces vs. rook or that a queen is worth more than two rooks in most positions (for example due to tactics) - but it is surprising how similar rating ranges can convert. I like how objective and clean you present the topic. You have my full respect. I hope to see more blogs from you in the future! Maybe you want to leave me feedback on mine too as I would value your opinion. Please message me if you create a new blog :)