A possible statistical parameter

in Ewn games is the average number of

“captures of enemy stones”, “captures

of own stones” and the ratio of these

two values.

A preliminary study (with a rather

weak computer program) shows that the

ratio (for board of size 5x5) is near

to 1.

My request to the writers of Ewn computer

programs: Can some of you run series of test

games (with strong level of play) to learn

more about the ratio (and its limit for playing

strength –> infinity)?

Thanks in advance for your help.

Ingo.

There's a reasonable corpus of games on LG that you could use.

e.g. all my games can be found here

It shouldn't take too long to write a program that simply parses those game dumps, calculates the tallies, and produces a result. For human players it's probably best to ignore their earlier games, as they probably didn't have a clue what they were doing initially.

Hi Phil,

thanks for your feedback.

> … For human players it's probably best to ignore

> their earlier games, as they probably didn't have

> a clue what they were doing initially.

This problem of fluctuations or changes in playing strength

is indeed a big problem. Therefore, my request was primarily

for automatically generated sequences of autoplay games of

computer programs.

Greetings, Ingo

I've just written a script which will permit me to play my bot at different skill levels head to head against itself.

E.g. 10 games at skill level 8 versus skill level 12:

phil@made:EinStein$ time ./h2h.plHits: Blue(8) hit 1 blues, 1 reds; Red(12) hit 3 blues, 3 reds. Winner=redHits: Blue(8) hit 0 blues, 2 reds; Red(12) hit 1 blues, 1 reds. Winner=blueHits: Blue(8) hit 2 blues, 1 reds; Red(12) hit 2 blues, 2 reds. Winner=blueHits: Blue(8) hit 2 blues, 2 reds; Red(12) hit 1 blues, 2 reds. Winner=blueHits: Blue(8) hit 2 blues, 2 reds; Red(12) hit 1 blues, 1 reds. Winner=redHits: Blue(8) hit 1 blues, 4 reds; Red(12) hit 3 blues, 1 reds. Winner=blueHits: Blue(8) hit 2 blues, 2 reds; Red(12) hit 1 blues, 2 reds. Winner=blueHits: Blue(8) hit 1 blues, 2 reds; Red(12) hit 1 blues, 1 reds. Winner=blueHits: Blue(8) hit 2 blues, 3 reds; Red(12) hit 1 blues, 1 reds. Winner=redHits: Blue(8) hit 2 blues, 0 reds; Red(12) hit 1 blues, 3 reds. Winner=blueTotal: Blue(8) hit 15 blues, 19 reds; Red(12) hit 15 blues, 17 redsTotal: 7 blue wins, 3 red winsreal    3m6.063suser    3m2.983ssys     0m0.316s

Is that useful (in bulk, that is, not just 10 games)?

> This problem of fluctuations or changes in playing strength

> is indeed a big problem. Therefore, my request was primarily

> for automatically generated sequences of autoplay games of

> computer programs.

Is it really a problem? It is a valuable resource, I think: analyzing the games of a player over time and observing the capture ratios. This could be connected to the rating chart for some informations how playing strength and captures correlate (the bigger problem is that opponents at LG-games are not randomly choosen and the avaluation would need the opponent strengths which would lead to a incremental process to get values for all players).

Question: is there a (proofeable) optimal number of captures in EWN (over many games, with optimal play, etc…) and are the capture ratios of a player a good estimate of playing strength?

> Question: is there a (proofeable) optimal number of captures in EWN

> (over many games, with optimal play, etc…) and are the capture ratios

> of a player a good estimate of playing strength?

case 1:

In theory you can calculate statistics across both-sides optimal play for all beginning positions. The result will be one number containing the truth.

case 2:

Things are not so clear if you let the optimal player play against non-optimal players. The (e.g. evasive) behaviour of the weaker players could have an effect on the capture ratio of the optimal player.

Evidently a weak player can have the same capture ratio as an optimal player.

Assuming you meant case 2 the answer to both questions is NO.

To Phil:

Yes, your “construction” is really useful.

However, in my eyes it would be most interesting

when you let rum a whole tournament, say

10,000 games for “each” level pairing:

level 8 vs level 8

level 8 vs level 12

level 12 vs level 8

level 12 vs level 12

(By the way: what has level with search depth to do

in your program?)

Thanks in advance for your test series!

Ingo

search depth = [(level + number of taken pieces)/2]

So level 10, where RoRoRo is currently, starts off as 5 ply. When two pieces are taken, it will kick up to 6 ply; 4 pieces taken moves it to 7 ply; etc.

I think there might be some bugs with either my script or with RoRoRo itself as far too often the lower level wins! It looks like it might be blue/red dependent too.

Hmmm, I don't like bugs…