Home
Information
Login
Register
Waiting lists
Membership
Hall of fame
Tournaments
Best game
Wikichess
Rating list
Problems
Forum
Links
Help
About
| Install FICGS apps | |
| play chess online |  |
|
Game result (chess) J. Dyson, 2044 G. Cesbron, 2208 1/2-1/2 See game 151864 |
Hot news Discussions Files search Social network
Here are 100 results for 1. Nf3 Nf6 in the games.
Game_152359 Game_152298 Game_152147 Game_152097 Game_152046 Game_152039 Game_152027 Game_151944 Game_151916 Game_151913 Game_151866 Game_151805 Game_151796 Game_151756 Game_151678 Game_151527 Game_151523 Game_151469 Game_151465 Game_151438 Game_151396 Game_151279 Game_151274 Game_151259 Game_151255 Game_151248 Game_151213 Game_151208 Game_151184 Game_151031 Game_150776 Game_150750 Game_150747 Game_150737 Game_150727 Game_150721 Game_150581 Game_150552 Game_150501 Game_150395 Game_150385 Game_150259 Game_150247 Game_150134 Game_150132 Game_150033 Game_150031 Game_150020 Game_149991 Game_149979 Game_149814 Game_149796 Game_149336 Game_149323 Game_149310 Game_149299 Game_149243 Game_149217 Game_149204 Game_149024 Game_148824 Game_148799 Game_148621 Game_148552 Game_148500 Game_148472 Game_148415 Game_148414 Game_148409 Game_148355 Game_148327 Game_148299 Game_148269 Game_148250 Game_148248 Game_148087 Game_147941 Game_147724 Game_147680 Game_147604 Game_147597 Game_147507 Game_147482 Game_147420 Game_147364 Game_147260 Game_147229 Game_147138 Game_147122 Game_147117 Game_147044 Game_147032 Game_146990 Game_146856 Game_146848 Game_146654 Game_146628 Game_146604 Game_146530 Game_146515
There are at least 1 results for 1. Nf3 Nf6 in the forum.
Attila Ba (2013-05-15 17:41:52)
Deep analysis - can it be improved?
The idea behind deep analysis is to store engine evaluations of chess positions in a permanent way and build an analysis tree out of them. Deep analysis is an improvement over simple engine analysis in two ways:
1) Permanent storage of analysis results makes them reusable. You don't have to analyse the same position from scratch over and over again (which is a waste of valuable CPU resources) rather you can build and improve upon your earlier results.
2) The search is configurable. You have control over which positions are examined and in what way. This gives you freedom to tailor the analyis to your own needs not having to rely on the defaults provided by your engine.
This idea is presented in a revolutionary way in the Deep Rybka Aquarium GUI. However using this framework I have encountered some problems. The lesser one and non lethal one is that draws by repetition are not handled correctly. This is for a reason: moves in the transposition table should be valued in an absolute way (regardless of the line which lead to them) in order to preserve the integrity of the tree. Since Aquarium has no means to incorporate lines, it simply ignores them
My other problem is that though the search is configurable I'm not absolutely certain about what is going on. It is not entirely clear to me exactly which nodes are selected for analysis.
These problems made me to try to come up with a deep analyis program of my own. After several failed attempts finally I have on my hand a solution which is not only capable of performing deep analysis but overcomes some of the difficulties of Interactive Deep Analyis (IDeA) provided by the Aquarium framework.
First I introduced a mechanism that can handle repetitions. In order to achieve this I attribute not one but two scores to each move and re-define the concept of root position already present in IdeA. The first score which I call 'idea' score is the same as presented in IdeA. The second is what I call 'alpha' score is calculated by minimaxing the tree from the root position taking into account repetitions.
Consider the following game:
1. Nf3 Nf6 2. Ng1 Nf8
The value of move 2. ... Nf8 at depth 18 by Houdini 3.0 is -19 centipawns. So the idea score of this move at depth 0 should be -19. Yet 2. ... Nf8 repeats the starting position. Therefore its alpha score with respect to a root equaling the starting position should be 0 centipawn which is exactly what my program calculates for it. ( For the sake of simplicity I don't require threefold repetition, since you would never allow your opponent to repeat a position if you have better ideas. )
So when my programs lists the tree it will present both scores for every move (which in most of the cases are equal of course - therefore this is mostly an aesthetic improvement rather than being a substantial one).
The improvement which I'm most interested in is that having full control of node selection now I have freedom to shape the tree search.
In order to keeps things simple I have only three parameters characterising the search:
1) engine depth
2) move distance (centipawns)
3) search depth
Engine depth means a fixed depth at which each move is analyzed. After long experimenting I have arrived at depth 18 as a good default for Houdini 3.0.
Move distance is a tolerance up to which moves are allowed into the analyis. For each position first the best move is determined. The search for alternative moves is continued until a move is found that has a valuation less than the valuation of the best move by 'move distance' centipawns (it is this 'distance' away from being the best move). The tree is then expanded for moves within 'move distance'.
To compensate for exponential growth of analyzed nodes I use a simple technique: at each ply after ply 1 the move distance is halved. So if the move distance at ply 0 and ply 1 is 20 centipawns, it will be 10 centipawns for ply 2, 5 centipawns for ply 3 and so on. This means that at greater depth less and less moves are allowed per position. So the analysis with greater depth slowly evolves into 'autoplay' rather than 'tree search'.
The other method to reduce exponential growth is the well known beta cut provided by alphabeta search. In order that all candidate moves in the root position and all candidate responses to them get proper values, I only allow beta cuts with ply 2 and deeper.
Once an alphabeta search of certain depth is carried out, the whole tree is mimimaxed out for the root. Now the initial evaluations of the root moves may change. This may make moves which initially fall out of the 'move distance' to become viable. So the search has to be repeated for those moves as well. This has to be done at every ply level.
My iterative search at a certain depth only ends when no new nodes are added by the alphabeta search (the tree is 'settled' for this depth). Only then the program is allowed to deepen the search (this I call 'refined' search).
With engine depth of 18 and move distance of 10 centipawns an average position can be analyzed to depth 10 within a matter of hours. This means a couple of hundred (possibly a couple of thousand) positions are analyzed to depth 18. Depth 10 deep analyis means an ultimate depth of 28 if you take into account that the engine depth is 18.
Whether this method has added ELO value over simple engine search is yet to be tested.
There are 1 results for 1. Nf3 Nf6 in wikichess.
Thibault de Vassal (2407)
Nf3 Nf6
Réti opening.
============
Contributors : Thibault de Vassal
[Chess forum]
[Rating lists]
[Countries]
[Chess openings]
[Legal informations]
[Contact]
[Social network] [Hot news] [Discussions] [Seo forums] [Meet people] [Directory]
[Social network] [Hot news] [Discussions] [Seo forums] [Meet people] [Directory]
Social network : create your photo albums, discuss with your friends...
Hot news & buzz : discover the latest news and buzz on the internet...
Discussions : questions and answers, forums on almost everything...
Seo forums : search engines optimisation forums, web directory...
Play the strongest international correspondence chess players !
|
Riccio, Eros (ITA) [member # 186] Correspondence chess : 2475 GM Ranked # 38 in the rating list. Rating history : No change this year. Advanced chess : 2443  This member may have a profile in the social network. |
 | Correspondence chess World championship Play chess games Go (weiqi, baduk) Advanced chess Play big chess Chess trainer apk Rated tournaments Poker texas hold'em Fischer random chess |  |