Reinfeld wrote:Hello, friends. Sorry for butting in and being absent. Great thread.
I agree that Igor/GM plays a strange but strong game. I am trying to find an analogy to describe it, and I don't have moves in hand to support my position; these are just impressions, without discussing the Nelson/Horvath question.
We have lots of measurements, but it's tough to measure feel, as in the way it feels to sit down and play these machines in a screw-around way, perhaps armed with a cooling beverage, hanging with Fernando and Thorsten.
1. Igor/GM is crudely materialistic. In this sense, it does seem to differ from the Horvath program in Systema Challenge, a more passive player to my eye.
Another way to say it - Challenge feels wimpy. It plays not to lose. Igor/GM doesn't do this. It swings wild, tries to counter and punch you in the mouth, but it recovers from misses. Igor seems to have different priorities - more like aggressive, anti-positional liquidation, but without vision.
2. Igor/GM doesn't play the pleasingly silly, speculative moves I associate with Kittinger machines. Nor does it have the steamrolling, pitiless quality of Spracklen machines, or the whiplash tricks of Morsch. It shows nothing like the anti-provocative style of Lang machines, and it is not well-rounded like Schroder, who is beginning to be my favorite programmer.
3. In style, Igor/GM reminds me most of another crude program I've never seen mentioned here: The Oak Systems engine for Kindle, a pretty simple player, recently jacked up with an opening book. This software is a dumb brute, especially at lower levels. It almost always takes material when offered, at the expense of position. Igor feels more sophisticated, but the approach is similar.
Now, despite my promise, an attempt to address the Horvath theory. Random thoughts, no particular order:
a. I've weighed in on this lingering debate elsewhere, noticing housing and level tuning similarities between Horvath/Challenge/Regency and Igor, but I think Nick's old argument suggesting that Igor is a re-tuned Horvath with more aggressive priorities makes a certain amount of sense.
b. What one wants to know is similarity, right? isn't that the point? How to find out? How to be sure?
c. We have pure hardware tests based on non-book moves (for instance, Steve, long ago, showed me the way to measure which model of Excellence I owned by starting with 1. f3 and measuring response time.)
d. We have Nick's ever-evolving clone test, which sets a threshold of 90 percent similarity while demonstrating conclusively that known machines of similar strength by different programmers indeed behave differently.
e. We have test suites based on positions. When I look at Wiki-Elo, I see no Colditz test results for Igor/GM. I see a set for Challenge/Regency/Legend II, however. Surely running a Colditz test on Igor/GM isn't too much trouble? No BT test results, either, though I recognize those are better for stronger machines.
f. The strong Excalibur machines add to that voluminous, vague ELO level of 1750-1800+, where some of our most beloved machines reside. And they all have the same tedious tendency to suck at endgames, where the horizon gets too long.
g. Is there an unexplored testing possibility here? A group of endgames beyond the horizon effect that would show the differing tendencies of the Excalibur machines (and others?) In other words, what is the choice these machines make when no clear choice exists? Any juice here?
- R.
I just realized a simple explanation to explain this GM enigma, this strength associated with old materialistic approach, etc. I think that is what happens when simple full width search is capable of going beyond the limit of 5 ply that was the average for middle game in old challenger by Nelson and others.
Even more; those old machines did not performed full wodth search. Many many years ago I asked the people of Fidelity about his searching techniques for Chess Challenger 7 and to my surprise I received an answer from the engineers -perhaps even from Nelson- with a diagram about that. I recall it vaguely, but by example in level 5 or 7, the scheme was full-full-5-3-2 or something of the sort, meaning full width just in the first two or three ply and then prunning until reachin 5 ot 6 ply.
With a lot faster processor and lot more memory GM can do brute force search many times more effective than that, more than 2 or 3 ply, perhaps 5 or 6, which in the old Zylog 80 processor with almost no RAM for some hash tables was not posible. So you are tactically lot more strong BUT still you exude the same air of the old models with lesser power because the evaluation still is centered in materialistic values.
Fern