World Draughts Forum

What on earth is 'TT-size = 16'? That means it used a hash-table of 65 KB, does not seem very appropriate with 16 threads sharing the hash-table.

That version of the engine does not know 'Bucket-size', it should be 'Bucket-shift = 2', it clearly was in the ini-file when I sent you this engine ...

I agree with Joost here. 2030 is way to pesimistic.

OpenAI intends to have a automated AI researcher in 2027. Surely it can write a simple draughts-program far before that.

In fact, i think you can do it right now.

What will change is that if you do it now, you may need several days of prompting ...

I just asked openAI codex to optimise the speed. It thought and experimented for 10 minutes, and it increased speed by 11% for the move generator.

Here is the newest perft for dragon, on an intel i9-13900KF, running at 5.4 Ghz. Nps in milions/second.



bb bb bb bb bb
bb bb bb bb bb
bb bb bb bb ...

> The program is recording the moves incorrectly during the game.
I will fix this next version

> I get the impression (I may be wrong) that the measurement is roughly the same as if I were using 16 threads.
16 or 32 threads will not matter that much. Depending on your processor, it is probably ...

Hi Krzysztof,

I will have a look at the issues you have.

Note that dragon 4.7 is not stronger than the previous version. And you still can use the old databases.

Michel

It has been a while, but Dragon version 4.7.0 is now available:


New look, new icons
Removed all references to paid version/registration. It is all free now.
Solved crash if you used 32 or more processor cores (engine is unchanged otherwise)
You can now download the 6 piece endings from view ...

Hi Michel,

I understand but I do not see clearly how the tempo_difference can be estimated by the patterns.
Let's take a simple example. Suppose you are dealing only with classical game. With a separate feature and by dividing the tempo_difference in for example 20 possible values from -10 to +9 ...

Hi Michel,

It seems I have a different view (but I have not your experience in reinforcement learning !)

I agree with you that the tempo_difference can be estimated by the patterns alone but does that mean that we should avoid to use such fixed feature?
My view is the following : because the ...

Hi Michel,

Concerning tempo difference do you also consider the patterns themselves can do the count or do you think that this feature is irrelevant?

I think the question is whether you can estimate the tempo difference by using patterns alone. The answer is a definite yes here. In fact, you ...

Hi Michel,

I never tried that experiment, but it is not intuitive to me that it would only be a minor disadvantage. Without an eval term for man difference, I would expect the engine might toss a piece just to gain some minor positional advantage. Do you understand why this doesn't seem to be a ...

Oh, I thought there were better index functions. Better = less memory :P

I am using this for checkers as well and initial resullts were quite promising. May I ask which optimization algorithm u use for the logistic regression ? I tried stochastic graidient descent which works but is still rather ...

In order to solve the problem you mentionned I implemented two points:
1) I initialised the value of my 20 features to a value near from +2000
2) As soon as I want to modify a value I take into account a handwritten constraint saying that these values have to be ordered.

Concerning the fixed ...

Hi Michel,

Maybe I am wrong but it appears to me that by calculating your f function you consider that the piece-difference is associated only one feature. Is it true?
FYI in my implementation each piece-difference value (from 1 to 20) is associated to one feature among 20. That way the learning ...

I don't think there is any 'human' meaning to the function. It just maps -infinite to +infinite to the [0,1] range, and does so with some particular nice properties.

One thing to consider is that when for example you are 5 pieces ahead and say f=2.5, winning another piece doesn't increase the ...

Hi Gérard,

The learning that i implemented, uses the logistic function to map the sum of features to a winning rate:

winrate = 1/(1+exp(-f))

The winrate is the number of points you can expect to gain from the position (loss=0, draw=0.5, win=1.0), f is the sum of the feature weights. The learner ...