AMD Ryzen

[BertTuyt]

Michel, you are right that with short time scales, the non draw rate increases.
But with normal tournament conditions, the programs who utilize all processing power will draw more and more.
I guess I could win against Bolt running the 100m, when his legs are tied, but under normal conditions , no way.

So I was referring to killer draughts, or breakthrough draughts.
Opponent modelling is interesting if you play weakers players/programs, but does not work against perfection.

Bert

[MichelG]

Michel, you are right that with short time scales, the non draw rate increases.
But with normal tournament conditions, the programs who utilize all processing power will draw more and more.
I guess I could win against Bolt running the 100m, when his legs are tied, but under normal conditions , no way.
Bert

I recently tested to see if scan is still capable of losing.

To test this, i played against single-core scan, with scan only 1 minute per game.

To give dragon some advantage, i gave dragon 6 cores. Due to parallelism limits, it then has about 3 times the amount of search capability.

1-minute-1-core scan vs 1-minute-6-core dragon: -35+18=447 (from scan's perspective)
In this unfair match, scan is losing 7% (+-1.2%) of the games.

To test a bit further i gave dragon more time:
1-minute-1-core scan vs 50-minute-6-core dragon: -17+0=43 (from scan's perspective)
With dragon having about 150 times as much processing capacity, scan is losing 28% (+-7%) of the games.

This means there is some improvement still possible, at least at very fast games.

Michel

[Fabien Letouzey]

So I was referring to killer draughts, or breakthrough draughts

This summer I will release a Scan version that (also) plays Killer and breakthrough draughts, so programmers will be able to do their own testing.

The BT implementation is naive, and optimised approaches are probably superior. As I understand, both MC and AB + patterns are used in the original (non-draughts) BT game. In Killer, however, Scan will be the real deal. Since the game is much less forgiving, the Elo gap between programs is 3x larger. Interestingly, I expect Wilo to be comparable between variants.

Fabien.

[MichelG]

One more datapoint for the draw discussion....

Latest match dragon vs scan 2.0 gave +22 -25 =941 (from dragon's perspective), at 1 minute/game, 6 cores. That's 95% draw.

I used 3-ply ballots for this match. One thing i noted that not all openings are balanced. For example after
1. 33-28 17-21 28-23 white is in a very poor position. Maybe black even has a theoretical win

After 1. 34-29 20-25 29-23 dragons sees a +1 score at 60 seconds/29 ply

Next up for dragon: i will do some experiments with machine learning for draw prevention against weaker players.

Michel

[Maurits Meijer]

After 1. 34-29 20-25 29-23 dragons sees a +1 score at 60 seconds/29 ply

Piece 24 will fall. It's easy to see for humans. In this position you can cancel out the attackers vs. the defenders on the same row.
You'll end up with a position like:

White has no defense against 4-9-14-19

[BertTuyt]

Just curiosity.
Is someone considering to buy a Ryzen (8 core) or Threadripper (16 core) system?


Bert

[Luzimar]

Just curiosity.
Is someone considering to buy a Ryzen (8 core) or Threadripper (16 core) system?


Bert

Hi Berty I do not know anymore I've been researching the I9 more here in Brazil it's still very expensive this costing 1,500.00 dollars