BB+ wrote:Chris Whittington wrote:The trick is to be adversarial some of the time (which usually involves a full on dismissal of the opposition case and a full-on assertion of one's own) [...]
This was not the working methodology of the Panel, nor particularly one that I find all that valuable. Maybe, as you say, I think too "academically/scientifically", but I think it has many advantages. For instance, the above confrontational model has frequently led to personal innuendo (in some cases quite shameless), which (to me at least) has been a major case of "irritation" that was absent in the Panel.
Chris Whittington wrote:By being non-adversarial for a couple of minutes, I was investigating with you to determine if the actual PST positions were actually as far apart as they perhaps seemed. Your reply is a bit too woffly to determine though. My guess is you would probably like to compromise but are worried about the loss of face that might ensue.
I could propose the opposite -- once I posted this direct challenge from the data, you "almost immediately" tried to bump the discussion to a meta-level...
As I say, I don't think such debating techniques are efficacious. Nor is guessing about my possible fear of humiliation if/when I am shown to be wrong.
The above challenge stands. As an example, I will investigate the (current) Stockfish numbers, which have been noted elsewhere to have at least some Fruit congruence, and see how easy it is to reproduce said numbers (and do so exactly, not just approximately) with the
Fruit 2.1 code. Again I don't want to pre-bias the discussion by naming a metric of "code changes" (e.g., if one can match 5 pieces with one change each, but kings are radically different and take 30 changes, is that worse than 15 changes evenly distributed? -- is changing a "ramping array element" the same as changing a "tuning parameter"?), but will report want I find, at least if this topic hasn't diverged to nonsense by the time I complete said task.
Here's a really "vague" description of a problem. Suppose the base PST "formula" for one PST is just +16 on the four center squares, +12 one square away, +8 2 squares away, and +4 on the edge. Suppose program A (say Fruit) uses just that as his PST values.
Program 2, an old Crafty or Cray Blitz uses those SAME values to initialize the board. It then biases the numbers to center the values around 0, roughly. So it uses 4 values of -6 -2 +2 and +6 rather than the 4, 8, 12, 16 in fruit. Then it adds in the value of the specific piece (say minor = 325) to every square. And then it does some root preprocessing and adds some values on top of that that try to attract pieces toward the enemy king. Say +10 for the squares adjacent to the current enemy king's position, +4 for squares one farther away, and +2 for squares one further (and that's it).
Now, do we call those "similar"? I would have to say "no". The original PST only contained a "distance from center" penalty, nothing more. The new PST adds the value of the particular piece (which can vary depending on endgame or middlegame phase, and it then adds more values to attract friendly pieces toward the enemy king.
One argument I am seeing from the "PSTs are very simple crowd" is that it is ok to take the above values, and say "OK, first we factor out that king tropism stuff, then we factor out the piece value stuff, then we normalize all the scores so that they start at 4, and we have fruit... IE what can we do to get from B to Fruit? And once they find enough, they say "OK, fruit is simple, this program is simple, they look the same. They ALL look the same. I claim that first, -6, -2, 2, 6 will play differently than 4, 8, 12, 16. I have already tested this. One can drop quite a few elo just mucking around with these values because it doesn't take much to over-value or under-value a piece with such changes.
SO. How to compare. If one piece of a PST initialization is the same, I say "what about the rest"? If they only have one piece of information in each, then yes, they are the same. If they are biased by some constant, then they are close, but not exact. If you superimpose other information like piece values, distance to opponent king, distance to passed pawns, value of squares that can't be attacked by pawns, then even though at a very elemental level they have something in common, they end up having more that is different and that would make them "not cousins."
The Rybka case is quite different. Here, the SAME code will produce either Rybka or Fruit values if you just change the individual term multiplies to correct for the approximate difference in material score differences. Yes, there are a couple of extra, or even missing (depending on your POV) bonuses for specific center squares, But if 60 are the same, or 62 or the same, and two squares were tweaked, that's still the same values...
