Hotspots

[Lyudmil Tsvetkov]

You seem to make a lot of posts and then you are the only one answering them over and over again.

Yeah, I try to anticipate people's reactions.

There are people who, even with a feedback, do not know they have been told something...

[Lyudmil Tsvetkov]

I do not see the point of a patch succeeding after more than 80 000 games; the book has already been spanned, and everything else is just external unwanted influence on the result.

If I had the chance, I would stop any patch after it spans the whole book and sprt is still not finished. Authors are there to find improvements to their patches.

However, if anyone ever pushes my last suggested values, with slightly increased queen value, I bet it will pass very fast.

If this patch passes, maybe it would not be too brazen to push a patch with changed queen values.

Pain in the eye.

After 80 000 games, so exactly the whole book spanned, it is at LLR 0.00.
What is the point of following another 80 000 games?
What would possibly make a difference, if all start positions have already been played?

Instead of following this for some tremednous length again and wasting valuable resources, I would have corrected the patch here and there, to see what happens after a fewer number of games, either it fails quickly, or it passes quickly.

But, as the sprt rule is such...

Considering your rook psqt patch, I hope it passes, but, failing that, I guess you might simply want to submit another patch duplicating the already gotten values from SPSA also for the eg. I do not see why anyone would mind.

Or, alternatively, you might want to reschedule another autotuning run, with 64 instead of 32 values, both for mg and eg, using the strt values for mg also for the eg. Then SPSA will tell you what to do, if ever.

[Lyudmil Tsvetkov]

Another funny thing about autotuners: you submit a tuning run, the tuner tells you what the values should be, you push a patch that fails sometimes, then you resubmit an additional tuning run, and this time the tuner already tells you the objective values have changed significantly.

Why so, when actually the term is the same? When you believe the tuner, the first or the second time?

Those tuners are just like humans, we also frequently change our mind.

[lantonov]

Pain in the eye.

After 80 000 games, so exactly the whole book spanned, it is at LLR 0.00.
What is the point of following another 80 000 games?
What would possibly make a difference, if all start positions have already been played?

Instead of following this for some tremednous length again and wasting valuable resources, I would have corrected the patch here and there, to see what happens after a fewer number of games, either it fails quickly, or it passes quickly.

But, as the sprt rule is such...

Not stopping a test is statistically well justified. It has to do with probabilities, and they depend only on the limits and not on the number of games played.

[Lyudmil Tsvetkov]

Pain in the eye.

After 80 000 games, so exactly the whole book spanned, it is at LLR 0.00.
What is the point of following another 80 000 games?
What would possibly make a difference, if all start positions have already been played?

Instead of following this for some tremednous length again and wasting valuable resources, I would have corrected the patch here and there, to see what happens after a fewer number of games, either it fails quickly, or it passes quickly.

But, as the sprt rule is such...

Not stopping a test is statistically well justified. It has to do with probabilities, and they depend only on the limits and not on the number of games played.

Here it depends on noise.

If there were 300 000 inique opening book positions, I understand running a patch till that number, but repeating the same lines does not make any sense for me.

[lantonov]

Another funny thing about autotuners: you submit a tuning run, the tuner tells you what the values should be, you push a patch that fails sometimes, then you resubmit an additional tuning run, and this time the tuner already tells you the objective values have changed significantly.

Why so, when actually the term is the same? When you believe the tuner, the first or the second time?

Those tuners are just like humans, we also frequently change our mind.

Tuning with tuners depends on parameters and starting values. If those are not suitable, you are measuring mostly noise, i.e. wayward results. If, on the second time, one chooses better initial values and parameters, tuning would be better and the values would be more real. However, SPSA is not the best method. I saw it with rook psqt values, most of which did not converge convincingly during the tuning.

[lantonov]

Here it depends on noise.

If there were 300 000 inique opening book positions, I understand running a patch till that number, but repeating the same lines does not make any sense for me.

Opening book positions and games are different things. If you look at the games, you will see that no two games are equal. The probability for this is virtually 0. Since the results depend mostly on the games and to a lower extent on opening positions (they play the same position both white and black), the games are what matters. You see that the gap increases with the number of games. With 80000 games the gap was 300 (to become green), and with more games the gap widens for pass and decreases for fail. This reflects the power of the test. With more games the test is more powerful i.e. it distinguishes a smaller difference. This becomes more complicated when including also the number of draws (draw rate) but it has all been calculated in terms of probabilities.
For our test I predict that we'll have a result before 100,000 games, more probably yellow.

[lantonov]

Yellow on close to 96000 games.

[Lyudmil Tsvetkov]

Yellow on close to 96000 games.

Yeah, it seems you are a fortune-teller.

Actually, I prefer it very much like that, instead of going for another 100 000 games.

Lyudmil, please push one last one with these values:

PawnValueMg = 194, PawnValueEg = 256,
KnightValueMg = 818, KnightValueEg = 846,
BishopValueMg = 840, BishopValueEg = 865,
RookValueMg = 1278, RookValueEg = 1294,
QueenValueMg = 2520 QueenValueEg = 2560

I know I last time told this would be the last one, but I expected it to fail quickly. It would be a pity not to do another try on piece values, especially when you hope it might perform better...

[lantonov]

I'll do it with low priority to wait for an empty queue.