> Maya experts have denied that 2012 is the end of the world
Indeed, it's just the end of the Age of the Jaguar, which happens every 25,920 years, and we've been here much longer
> Veteran Rajlichologists will, of course, interpret Vas's new Rybka 5 release date to be December 31, 2012, assuming he is going by the Julian calendar.
I think his year starts in October. Alternatively, we could interpret that Rybka will be either 3 months or 3 units (years in this case) late.
Any idea on how much Rybka 5 will increase in ELO.
Will Rybka 5 be out in the first part of 2012.
Thanks for any information on Rybka 5.
Thanks for your continued work on Rybka 5.
Release date - I'd guess that summertime is the most likely.
Lets keep it easier for you to answer.
Do you intend to release Rybka 5 stronger than the actual Houdini 2 ?
Which week depends on global weather patterns which week are we talking....
> Elo - I really don't know yet. It will depend on how things go and where other engines stand at release time.
> Release date - I'd guess that summertime is the most likely.
An interesting post! I take it from this:
a) private Rybka is stronger than anything out there
b) (slightly less powerful) Rybka 5 will therefore depend on how strong the opposition is.
Do you have an ETA for when the Cluster can be rented for shorter periods of time?
(from now on anything in italics is copied from Watkins' paper)
Namely, an evaluation function has two parts: what to evaluate, and how much weight to give
each feature. Again the ICGA process recognized that Rybka had indeed varied
the latter; the complaint was that Rajlich copied Fruit in the former.
1- What are the big differences between Rybka and Fruit regarding evaluation?.
I personally find it completely clear that Rybka 1.0 Beta used the PST framework of Fruit 2.1. One can recover the exact Rybka PSTs
from the Fruit 2.1 code with 17-18 changes of numbers (tuning, as before), and
changing a mere 4-6 lines of code (all but one being an adjustment of the score
for a specific piece-square). No other examined engine required such a small set of
changes (firstly even assuming that it had the same PST scheme). Notably, even
Fruit 1.0 differed significantly from Fruit 2.1 here.
2- Did you copy the PST framework from Fruit 2.1, and if not how do you explain the similarities?
Regarding feature overlap, Watkins writes: In his last paragraph Riis pretends there is an issue of \data interpretation" with
the raw percentage overlaps. The method to interpret these is given in the ICGA
documents. Namely one applies a (standard) statistical analysis to the raw percentages,
and measures the variance of the Rybka-Fruit pairing against the control
group. This should be fairly routine for anyone with a mathematical background,
and moreover is indicated in both EVAL COMP and the ICGA Report. Thus I find the query of Riis here to be rather odd, almost sciolistic in nature. At any rate, one concludes that the Rybka/Fruit data point was an outlier at 6-8 standard deviations
in a pool of around 30 comparisons, or more than a 1 in a million chance of occurring at random.
3- Would you comment on this argument please?
3.4. Rybka/Fruit copying that Riis omits. Here is a brief enumeration of a few
of the points not covered by Riis. As noted above, these are expanded somewhat
in Section 4 of the recapitulation PDF.
(i) The Rybka iterative deepening code is largely the same as Fruit's, adding
only a few elements (like mate scores), and varying a few parameters. The semantic similarity is high (both check a variety of conditions in the same order), and the six variables of interest are allocated identically. I don't
know of another engine that uses the general Fruit concept here, let alone
one with so many code congruences.
(ii) Rybka's root search follows the same pattern of that of Fruit, including more
than 5 notable points of commonality. An examination of other engines
showed none that followed said pattern.
(iii) Rybka copies a small piece of search control code from Fruit, where again
the ordering and variable allocation are the same. The Fruit code also includes
some idiosyncratic redundancies, which curiously re-appear in Rybka.
(iv) The hashing structure of Rybka is the same as Fruit's, except for transposing
bytes 8-11 with 12-15. Again I cannot nd another engine with
anything so similar; indeed, already Fruit 1.0 differs.
4- Did you as suggested by Watkins, produce an engine more similar to Fruit than its own predecessor Fruit 1.0?
4.1. PST. To pre-empt Riis (as he diverges rapidly), the evidence here is clear. As
I stated above, the Rybka 1.0 Beta PST can be generated from the Fruit 2.1 code
by changing 17-18 numbers (tuning), and 4-6 lines of code (amongst about 50). In
the hopes of defogging the morass through which Riis wades, I challenged anyone
to nd any other engine (not Fruit-related) whose PST could be derived from such
a small set of changes from the Fruit 2.1 code. There has been no answer.
5- Do you know of any other engine (not Fruit-related) whose PST's can be derived by a small set of changes form Fruit code?
What I can say is that Rybka is original at the level of source code.
For the super-geeks, yes, that can be applied recursively.
For other algorithms, it would show at the very least a heavy influence of one program on the other.
You have repeatedly claimed that ALL versions of Rybka were original, including the early versions that pre-dated Fruit. Can you offer any explanation for the findings of the ICGA panel regarding Rybka 1.4 thru 1.6.1 and Crafty 19.x? There are MANY identical sections of code in both. Some of the code was utterly useless in both because Edwards tablebases were not available after Nalimov came out. Some of the code was vestigial in Crafty, left over from previous versions. The infamous
if (ms == 99999) break;
Yet if you look at EvaluateMate, it can NEVER return 99999. At one point it could, but that slipped by me as I released 18.0 (it was an early approach that I changed for efficiency, but forgot to remove that test.) This is the same code that uncovered El Chinito as a Crafty clone, which was a thread on CCC that you once posted in.
There are many other examples. For example, code in iterate.c to avoid egtb probes in KPKP ONLY, when the two pawns were on adjacent files and one of them was on its original square. Steven't EGTBs did not consider en passant, so it was unsafe to probe if an ep capture was possible.
There are dozens of other identical things between the two programs.
So, exactly what gives? You entered these versions in CCT, which had the originality requirement. You submitted them to others (such as ChessWar) which also expected original programs. Can you explain that, as a first step. It looks really ugly.
I'm not going to go into this today. My view is that these are private versions which are my private business. This is something for another day.
Regarding Strelka/IPPOLIT: as the author(s) seem to have typed their own source code (or code to generate this), how are they not "original" under your definition? Do these Rybka versions [the ones that were cloned] have any additional creative content beyond the source code?
> Regarding Strelka/IPPOLIT: as the author(s) seem to have typed their own source code (or code to generate this), how are they not "original" under your definition?
I doubt that all of that code was typed by hand. If it was, then sure, it's "original at the source code level".
> Do these Rybka versions [the ones that were cloned] have any additional creative content beyond the source code?
Lots of brilliant ideas!
Thank you, Jeremy
At this rate we'll have a page-long definition by tomorrow night.
> I'm sorry, can you please clarify this? Does this mean, if I have source file A in one editor window, and my new source file B in another, and I type, by hand, the contents of A into B, that I've created software that's "original at the source code level"?
> Thank you, Jeremy
Can you please clarify this ? If I type some code, and it is semantically equivalent to a piece of code in someone else's software, am I guilty of code copying even though I have never seen that other program, or even be aware that it exists ?
I don't really know what this has to do with Vas and Fruit, though, as he certainly had seen that other program, and went forward and backward through it, taking many things.
> No, of course not.
Bo Hyatt would disagree, he says this is code copying. He stresses time and time again that semantic equiivalence is enoguh to prove code copying.
But only if Vas is doing it, not if he is doing it by himself.
And one more curious question.
Do you consider that Houdini is a clone (or a derived clone)? Do you think Houdini is an original engine that doesnt have any rybka code?
Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill