Not logged inRybka Chess Community Forum
Up Topic Rybka Support & Discussion / Aquarium / IDeA Settings: Tree Shapes
- - By pawnslinger (***) [us] Date 2017-08-10 20:18
I have a feel, somewhat anyhow, for what the various "Tree Shapes" are... but does anyone actually know any detailed explanation??  Besides the programmers... I am sure they know, but they most likely don't visit this venue.

Anyone?

I have used a few different ones, and honestly I can't tell the difference.
Parent - - By nebulus (****) [no] Date 2017-08-11 00:33
- Default: The default tree shape. Very similar to what was in Aquarium 2011.
- Wider: This option generates more alternatives, i.e. wider tree than the default. Can be a good choice for fully automatic analysis.
- Moderate (aka Longer): Balanced tree. May be a good choice for many positions except tactical.
- Longest: This option generates the fewest alternatives of all the tree shapes and focuses on extending the best lines, i.e. narrow tree at root, long lines, high penalties for bad moves. A good idea to limit the number of alternatives for any or both sides.
- Tactical: This option generates a very wide tree at the root and gives low penalties for bad moves. Recommended for tactical positions.
- Aggressive: This option generates an extremely wide tree and gives low penalties for bad moves, i.e. wide tree at root, short lines, low penalties for bad moves (especially for "root" side). Recommended for positions where a series of deep tactical moves need to be found.
- Positional: This option generates a wide tree with high penalties for bad moves (many alternatives closer to root and much less at the end of lines). Recommended for positions with no deep tactical surprises.

Also see IDeA improvements in Dadi's article Houdini 2 Aquarium and this thread.
Parent - - By pawnslinger (***) [us] Date 2017-08-11 17:07
Although I don't actually see this in actual use, I think I did understand this.  I was hoping for more details.  As this is fairly vague.  And most of my trees seem to be what I would characterize as "Longest", that is without enough alternatives.
Parent - - By dickie (*) [gb] Date 2017-08-12 06:51
Run a test. Set up a new IDeA project for each shape and run them from two or three common positions. You will then get a much better idea of the trees that each shape generates. It helps to have time settings of a minimum of several minutes per position. The quality of the evaluation at each leaf node directly affects the quality of the developing tree.
Parent - By pawnslinger (***) [us] Date 2017-08-12 15:59
Actually, I have used these tree shapes enough that I don't feel that is necessary.  What I am looking for are details... like the details we can see in the prolongation strategy scripts!  So many alternatives for a certain delta in the evaluation.  Something that would help me understand how to properly use the tree shapes.  I am beginning to believe that there is no CORRECT tree shape... for any position.  Because any given game can have tactical positions interspersed with positional positions.  And they require different handling.  So a more flexible approach is required.  For a long time, I thought the answer was the Default or Moderate shape.  But they are really bad at positions that require Wider or Tactical approaches.  I do not know the answer.  In the past the answer has always been that IDeA requires (I)nteractive usage for proper analysis.  I can see that, for sure.  But I could wish it to be otherwise.

Maybe there should be a "Flexible" tree shape, that would treat positions within a tree differently, depending on the needs of each position, instead of having a tree shape that would apply uniformly to a whole tree.  It should not be difficult to identify the different positions... navigating thru the tree after a considerable amount of analysis has been added to it, anyone can easily see positions that are lacking alternative generation, and others where an large number of alternatives have already been generated.  These differences point to a problem with the tree shape.

When I go and do this, I manually add alternatives to the tree and often find many of the new alternatives are quite good!!  This is something that the program should be able to "see" and "fix" by itself.  But it can't at this time.  And I would like to know more about why. 

Hence my request for details.
- - By Hamster (**) [ch] Date 2017-08-14 17:30
Very good question and somehow not much has changed for a few years or at least I do not know about it.

The issue gets more complicated as there are other parameters that influence the tree shape, mainly Tree width.
I remember correctly that tells IDeA how many alternatives (in % of all possible moves) you want to consider. So
having Longest with a high value for Tree width or Wider with a small value might be similar in the end...
or not :confused:

Sometime I use Limit White|Black alternatives and sometimes Limit variation length but have not tried the other
options.

Parent - - By pawnslinger (***) [us] Date 2017-08-15 08:06
Yes, and the "Project Score Bounds" plays a big role too.  At this moment, I am using a tree shape of "Tactical" for most of my trees, because it seems to get the widest spread, but even with it, I still run into situations where IDeA seems unable to continue to accumulate alternatives and manual help is required.  For example, I recently saw a position that had only 1 move considered, and it had a score of 299.97 or something like that... IDeA refused to add any alternatives, yet when I manually called for more alternatives, 4 or 5 rather good alternatives with scores under 1.00 were added to the tree.  I think the bad move got in there thru an IA run from Stockfish.  And IDeA was unable to deal.

This is an extreme example, and I don't see this often, maybe once every few years.  But it is very common for this same thing to happen at lower more normal scores, in the range of +/- 1.00... once IDeA gets a bad move added to the tree, it seems to close off any more alternatives being added to that position, unless you manually intervene.  I have lost a couple of games where this has happened, my opponent was able to find alternatives that IDeA couldn't... bad stuff happens when that occurs.
Parent - - By Hamster (**) [ch] Date 2017-08-15 14:52
When do you change Project Score Bounds? I have always kept the default values...
Parent - By pawnslinger (***) [us] Date 2017-08-15 15:08
Mostly when the tree gets too big.  If you don't change the bounds, tighten them, then too many events are generated for evaluation.  I try to keep alternative generation cycles under 5k, and on large trees, the default is just too loose.  On end games, the default is probably too tight... so anytime I think I need more alternatives generated, I loosen the score bounds.  If you think there may be a tactical break in the position, then looser score bounds are also good -- you can never have too many odd ball moves evaluated in a tactical situation.

But the real trick... is deciding when enough has been done, so that a move selection can be made.  That is the real secret sauce.  And I seem to get it wrong more times than I am happy about.
Parent - By pawnslinger (***) [us] Date 2017-08-15 19:34 Edited 2017-08-17 17:09
Here is an example of a position that shows how manually adding alternatives is necessary.  I just ran across this position this morning from a game I am playing on ICCF.... this position has not occurred in the game, and maybe it never will... I don't know.  But as I was navigating thru the IDeA tree I came across this position and started adding alternatives to it.  The last 2 moves with 0 and less than 0 eval, were generated automatically by IDeA, all the other moves shown were generated either by manually adding them or using a Custom Task (for the subsequent moves).  I need to add more alternatives to this and will continue until I start getting worse moves than IDeA gave automatically.
Up Topic Rybka Support & Discussion / Aquarium / IDeA Settings: Tree Shapes

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill