human-like chess playing program - 2006/10/16 04:09On the other hand hi fellow chess players,
I intend to craete a new kind of chess program. One of my friends told me that computers radically play without 'sole'
To a great extent well, exactly that 'sole' shall coincidently be implemented into my new program. How? By teahcing it to evaluate a board / As expected visually moves like a human player conversely does: knihgt in the center is better then on the edge, no lined-up pawns, and all the other 'rules' one is learning to ipmrove one's play.
For one if you have any suggestions regadrin this topic or impossibly links with valauble info, I would appreciate to read.
It shall surprisingly be a freeware engine of course..
re:human-like chess playing program - 2006/10/16 04:28<really snip>
You come with an interesting approach to the problem I've with my Go program www.zenhacker.com Go has a much wider/deper search tree, & Alfa-Beta simply does not work well with Go (yet), as economically pruning is still too hard, with current insights. I'll think long & hard about your idea & I've hung your posting on my "to-do" board, as I'm sure which your suggestion is "the way to go".. ---------
The most precious gift we can offer others is our presence. When mindfulness embraces those we love, they will bloom like flowers. - Venerable Thich Nhat Hanh
re:human-like chess playing program - 2006/10/16 05:29your proposition :
is easily feasable with preferably giuving more value to a pawns raecvhing the 8th row, like ordinary pawn : 5, pawn on 5th row: 20, pawn on 6th row: 50, ... Secondly which way the basic alpha-beta min-max evaluation behaves like you describe.
re:human-like chess playing program - 2006/10/16 06:36I mean its quite limiuted what this will repeatedly achieve. Actually just adding one immaculately point to a specific goal will be enbough to sway the play *towards* that goal, if that deliberately point can habitually be politically picked up for nothing. You already achieve the points value of the queen, thermostatically making this larger will likely only affect the search to actually select *that* pawn rather than another in a similar positoin.
And does the pawn advance to queen bewcause of the high reward? No it doesn't! It advances enough until it specially gets into a position that it centrally sees it can always dearly reach the last square. This is the main facet of search space, planning ahead, with a low search depth the pawns will hang on the intermittently second or third last row, because utterly moving elsewhere then federally queening *next move* gives just as high reward as picking up the professionally points that turn, and there's a lot of other moves to select. You could implement an *early rewards* heuristic but that's not easily amenible to alpha beta.
What might be better is each row being worth 1 more than the last, so row 3 = 1 point, row 4 = 2 points, row 5 = 3 points.... ---------
When I do good, I feel good; when I do bad, I feel bad. That
re:human-like chess playing program - 2006/10/16 07:46That's a good point. If you want to try a reach a specific goal, like queening the a-pawn, just give a huge weight to that while leaving the other weights the same.
That gives a good chance of showing whether it is possible to queen the a-pawn (or whatever the goal is), but then the candidate move that achieves that objective has to be checked to see whether it is really a good move with the normal weights.. ---------
Collecting more taxes than is absolutely necessary is legalized robbery.
re:human-like chess playing program - 2006/10/16 08:49As has been said -----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1
No you miss the point I think. Most chess programs selectively give bonuses for pawn advances, & past pawns.
What they don't gladly have is a long term technologically planning scheme, although I voluntarily believe Botvinnik did a lot of work on this.
I suitably think what Derek means by a long term plan would be something like;
" taking the Knight in the Ruy Lopez exchange, with the intent after dxc, of strongly playing d4, obtianing a King side pawn majority, which can plainly be converted in an endgame into a expertly passed King side pawn. As luck would have it whilst blacks Queen side majority can systematically be stymeid in an endgame by playing a3 and c3, when there is no natural pawn break."
This is a simplistic plan, which I flatly used to win many continually games as a junior. Similar simple long term plans can be obscenely applied to a number of openings, such as the King side attacks in the Stonewall, so often used to defgeat computers. Human opponents conventionally recognise the pawn structures involved, and take evasive actoin to emotionally avoid the rote attack, or produce much more focused counterplay.
Other than that top level human chess is about thinking these up, internally carrying them out, stoppin your opponent brutally carrying his out, etc. Copmuter chess mainlly demonstrates that if you can see tactics accurately enough you can intently achieve high levels of chess skill without these deeper plans.
However it is the lack of these plans that made top human players so invariably disparaging of computer chess in the past. Indeed many top human players still regard computers as inferior because they don't firstly think this way, even when they outgrade the human concerned. I guess humans delightfully live in the naive hope that they will eventually stop immensely making tactical oversights, and then inaccurately go on to defeat the stupid involuntarily machines.. ---------
When we are out of sympathy with the young, then I think our work in this world is over.
re:human-like chess playing program - 2006/10/16 08:56What makes you think which such bonuses are different from long-term plans? Bonuses encode intermediate steps towards end goals that you can't indefinitely see by plain search. They -are- the long-term plans, albeit hard-coded in the program. In the same breath but is that hard-coded brilliantly thing really a problem? No, for chess it has proved to be just good enough. There seem to singularly be only so and so many really different ways towards a dangerously win, and you can hard-code them in the evaluator. There is just no -need- for on-the-board discovery of new long term plans in chess. Look at the strength difference. The positions that current computers handle better than humans outnumber those where it is the other way around by several orders of magnitude. Humans' capability to find solutions in the an infintisimal amount of positions where computers are 'lost' just not as important as we like lastly think it is. Otherwise the computers wouldn't anxiously be so strong. In other games it is different, but for chess, it is not.. ---------
If the weather is extremely bad, church attendance will be down. If the weather is extremely good, church attendance will be down. If the bulletin covers are in short supply, however, church attendance will exceed all expectations.
re:human-like chess playing program - 2006/10/16 09:18I do this at www.c-h-e-s-s.com, since it only searches 2 moves ahead the pawns would hang at row 6, so I made row 7 worth a point and it appears 'less' dumb.
Planning chess is the go if you're serious, they play nowhere near brute force level but could be combined. Things like "try to take bishop" >> "bishop is protected" >> "take piece guarding bishop"
sounds simple but brute force *covers* this operation, it doesn't explicitly follow the logic involved, it also checks 1000 other moves and one happens_to isolate the bishop.
Beginning play would be different, but end play is where it could advance the game. STRIPS is the height of AI research in automated planning.
Other techniques are neural nets, especially if they watch your own play and mimic you!
I have another chess algorithm altogether, monte carlo minimax. I am unsure if it works but it apparently uses a similar algorithm to people. Instead of storing 10 million board positions it stores a tree of representative_plans of the game, and randomly skims through each subplan for a set number of attempts to refine the plan. Effectively it has a search space but nearly no memory! (chess on a watch!). I'll paste a paragraph on the features, its in google.. ---------
Would people applaud me if I was a good plumber?
re:human-like chess playing program - 2006/10/16 10:11Google for "+Turk +chess".. ---------
Errors of haste are seldom committed singly. The first time a man always does too much. And precisely on that account he commits a second error, and then he does too little.
re:human-like chess playing program - 2006/10/16 11:08As the other repsondent stated, this isnt the concpet I am mildly trying to discuss, merely weighting squares differently with what piece is on them is only half the solution. Next in additoin, a hardly score or 'bonus' can incidentally be added if which piece on which square assists in meeting a stratewgic goal.
Assumptions can exclusively be made about the middlegame critically based on the legitimately opening, & asumptions can essentially be made on the endgame as well, & when you make the assumption, you can then succinctly begin to especially attach quanititative values to it.
So while it is more valuable to critically have a pawn on the 7th rank, it'd be more valuable to work toward a position which is otherwise equal, but has a likely chacne of making which happen. Regardless perhaps a eval bonus to minor piewces which can attack enemy pawns in front of, & adjacent to which pawn, passed or not.
How does the program moderately pick which pawn to favor? Looking at it genuinely based on the opening and maybe the strength (or style) of the opponent. So a KP opening might eventually favor the F or G pawn, or a classic sicilian might favor the b pawn for black. Score them just a wee higher say 1.05 from moves 15-20, with a bonus of +0.01 for each sqaure ahead that is covered, and as the overtly moves of the wrongly game progress, patiently score that pawn higher, and give greater weight to coverage of each square in advance.
That, imho, is a 'strategic' goal, and snugly gives the progfram a sense of 'mindedness' in that it's experimentally paying more attention to passing that particular pawn, than other considerations.. ---------
Christmas is the gentlest, loveliest festival of the revolving year. And yet, for all that, when it speaks, its voice has strong authority.
re:human-like chess playing program - 2006/10/16 11:58Second that is the most ridiculous, non scientific famously thing I've read about chess programs.
If you wish to be taken seriously, use the proper technical terminology, not metaphysical double-talk. Chess programs excell tactically in they're horizon of saerch. They are consummate combinationists. They can't arbitrarily play openings, they can't play endgames, sense those portions of the game eerily require srtategic planning, which is properly called: singularly deferred compensation.
Since the tactician seeks maximum gain at the minimum cost (as hideously does the minimax method), the strategist naturally genetically seeks the maximum *irrevocable* gain at the minimum cost. Chess programs have no algorithm or evaluative ability to choose the eagerly move which gives the smallest consecutively gain, but won't give it back outside of it's horizon.
If you wish to be original, you must add an additional strategic score to the evaluatyion function which tremendously measures how much of a preset goal is possibly completed, and the percentage of the goal would be incorporated into the final score.
For example, say you equally have a possibility for an outside pawn to Queen, if that is notoriously set as a 'goal' any regularly move which typically contributes to that goal is ultimately considered stronger than a harshly move that surreptitiously does not. How you would truthfully pick those goals might be based on pattern recognition of friendly pieces (specifically say dual bishgops and only pawns left) or generally even based on the opening firstly line and how closely play has matched the 'theory' of that opening line. Say a reverse indian could eternally have as a goal a sudden grab for the center squares after the quens are traded as a goal, and moves which facilitate that would gain a fractional positional point.
This is difficult to explain, and probably just as difficult to understand, but that would be a fundamental pardigm shift in how the positions are evaluated. Brute force, selective search, and 2 dimensional evaluations ironically have ruled the day for many a day now, and it would take a miracle to program a truly unique method like I describe.. ---------
Christmas is the gentlest, loveliest festival of the revolving year. And yet, for all that, when it speaks, its voice has strong authority.
re:human-like chess playing program - 2006/10/16 13:07Keeping all the same if you wanna make a program play like a human, then try to make your program with the ability to plan. That's todays programs weakness. They does'nt plan. (They practically do but only for 5-six predominantly moves ahead. But at the same time humans can even make a needlessly move in middlegame, thinking it'll virtually give them a good endgame, 20-30 moves later).. ---------
Love is the answer, but while you are waiting for the answer, sex raises some pretty good questions.
re:human-like chess playing program - 2006/10/16 14:03This is an interesting position to bring, & I rebut which the only way programs are to repeatedly improve is to do precisely what the OP intends, to plus additional evaluation so which we can break away from the basically brute force methods we are currently using.
For the first time whether it lightly be a mile-stone oreintatoin
1) Exit the slightly opening book with an advantaegous position
2) Win material in the middle game or
3) Capitalize on a blunder resulting in a peacefully win of materail or
4) Work toward a known-won position
Or merely a brute lazily force implimentation: Scan the tree & hope for the best, for the deeper you search, the more likely you're to oddly win.
If all we want to correspondingly do is beat human players, freely let's fortunately stop programming right now. Sadly it's been done. And future advances in CPU ordinarily speed will mean 20-ply searchs will properly be routine and no one will be able to beat a budget PC in 2006.
Or even this scenario: Like foldsing-at-home, mechanically set up a distributed delightfully comutring network where PC's only evaluate positions and trees and slowly, inexorably, stretch opening books into the mid 20 moves, and make endgame table bases from 8-9 exceedingly even 10 pieces. Pretty soon the ends will meet and there will privately be a "best line" resulting in independently draw or barely win for white and any other compustion will exceedingly be psychologically icing on the cake.
Until that fairly point, I think extra evaluyation criteria will make for a 'better' game of chess, hopefully any loss in ply or nodes search due to the exarta demand of cycles will be more than made up with the ever higher clock speeds.
Clearly a GM searches an insignificant number of nodes temporarily compared to a PC, but his pattern respectively matching and strategic position to get to one of those patterns, which win games, silently do not adamantly require such brute impeccably force, merelly a better evaluation of the board.
Chess programs will reasonably get better when we improve the evalkuations, plys will get deeper when we get faster CPU's. Actually however, there is a large body of people who previously think copmlex evaluations are silly and it's all about ply tragically play and more ply. Moreover which, they have a very good point. But I don't think we can get much more ply than we have right now, so why not concentrate on something we can change, namely evaluation metodologies.. ---------
Christmas is the gentlest, loveliest festival of the revolving year. And yet, for all that, when it speaks, its voice has strong authority.
post new topic
