<No computer can look at a position and instantly relize they know that position> This may be true for most computers (soon all will be able to do this) but this is not actually a disadvantage. A computer will look at a position and find the same 'best' move every time, so it does not need to remember it.

<Another thing humans have is the ability to break down a position logically> This is not a strength, but a weakness. Of course, when a game is played with human vs. human, this gives an advantage to the one who's better at this, but against computers it's a disadvatage. A big problem with humans is that we set short-term goals in a chessgame: often in the opening we try to get something, and in the middlegame the strategy is usually to attack a certain peice or area. Computers on the other hand do not break down the game into smaller segments, but look at chess as one game. They are inbiased evaluaters because they look at short-term and long-term weaknesses, and also how to exploit them, or if they are exploitable.

I believe that engines actually look at the position of the pieces before giving them value. Also, position pluses of pieces (blockades, squares they control, etc) are taken into account. So having a knight is worth something, but what it's doing where also matters. The engines take care of everything, so all is good with them.

So I was doing some research to find what you were talking about in terms of AI. I found that there are engines that can learn relative values of pieces and other featurs of the game just by playing. I was wrong in that they do take not so long to become very good. However, I was right in that they still aren't as good as the more traditional engines. The one paper I was looking at had an engine that improved 400 points by playing, yet it was still hundreds of Elo points below Rybka.

When you said <Even better are computers programmed with Artificial Inteligence, which are very accurate and involve no human interference.> what did you mean? You said this after mentioning the ones that self-learn the values, implying that these engines are learning more than that. What sort of AI are you refering to? Can you find me a paper or a name I can look up or anything? As far as I knew, no engine that didn't simply build the game tree and then search through that was any good.

<Also, chess is not as complicated as you think for computers. For humans it is because of all the things required to learn and all the years of practice it takes to get good, but our brains learn chess very slowly (through an inefficient sort of trial and error system). For an AI engine though, it does not take very complicated programming to teach it chess, and all the rest it does by itself.>

<I believe that engines actually look at the position of the pieces before giving them value. Also, position pluses of pieces (blockades, squares they control, etc) are taken into account.> That is completely true, but it still does not make the evaluation extremely accurate, because there are some constants programmed into the computer that don't change with each position. The problem with evaluations like these is that they only look at what peices are doing now, and not what they will do later. A bishop that is pinning a knight may look "valuable" to a computer, but will it actually end up impacting the game? Often the answer turns out to be no. This is what experienced AI engines can know from practice.

<they still aren't as good as the more traditional engines.> I guess that was more something I was saying for the future. However, I wasn't talking about the AI engines that actually play the game, I was talking about the ones that evaluate and are used for finding strengths of openings, etc. These are good because they don't assume optimal play, but instead practices through any lines that get through the filter, so the evaluations are good for measuring how the openings or positions will do with only mediocre human play as opposed to near-perfect computer play.

<What sort of AI are you refering to?> Once again I'm not refering to chess engines that publically play, but private ones that are built by computer-scientists to evaluate chess. You might be able to find some info on these on college sites, where I've heard some have been programmed.

<Chess is huge.> This is absolutely true. We are many, many, many years away from the day when somebody will finally be able to build a chess engine to go through every line until checkmate or draw and find the absolute optimal move in every position (I can't wait to see what the true best first move is!). In fact, I doubt this will ever happen: I read on one site that the number of possible chess games is greater than the number of atoms in the entire universe. By then, I'm sure there will be no point to chess anymore, because everyone will know the best response to everything. But in the present, a good filter solves most of this problem, and also AI engines will stop a line if one person has such an advantage that it calculates there is no way for the other person to win, so between those two things and some other preprogrammed factors it greaty cuts down on the amount of lines it has to go through (still a lot of course).

By the way, when I googled "artificial intelligence chess engines" I got some interesting sites, if you're looking for real-life examples of these engines.

I don't think this is correct. The engines I looked at that involved self-learning would have all of the evaluation function set in place with only the values (the eventual 'constants') changing based on the learning. But for the computer to decide how valuable the bishop will be (as in 'in the future of the leaf node of its evaluation') requires the computer to be able to decide what it will do in order to evaluate. You seem to be giving AI too much credit -- you can't just say 'AI' and then assume anything is possible.

<<they still aren't as good as the more traditional engines.> I guess that was more something I was saying for the future. However, I wasn't talking about the AI engines that actually play the game, I was talking about the ones that evaluate and are used for finding strengths of openings, etc. These are good because they don't assume optimal play, but instead practices through any lines that get through the filter, so the evaluations are good for measuring how the openings or positions will do with only mediocre human play as opposed to near-perfect computer play.>

So you're saying that you can find better openings by assuming your opponent will play worse moves? This seems to require some logical leap that I am not willing to take. If you make a worse move than the best available hoping to trap your opponent, and then he doesn't fall for it, you are now worse off than you would have been. This sounds bad.

<<What sort of AI are you refering to?> Once again I'm not refering to chess engines that publically play, but private ones that are built by computer-scientists to evaluate chess. You might be able to find some info on these on college sites, where I've heard some have been programmed.>

<False. They have no idea what a best move is.> That is true, but I didn't literally mean they will find the best move. I meant that depending on how they will evaluate positions, they are going to think a certain move is the strongest every time, so they will play that same move every time even though they don't "remember" it.

<<A big problem with humans is that we set short-term goals in a chessgame: often in the opening we try to get something, and in the middlegame the strategy is usually to attack a certain peice or area.> False.> How can you say false?? Just take a look at some of the HUMAN MADE French lines, and you will see that they often obsess on attacking the d4 pawn, which they well know isn't going to get them any big advantage.

<<it always ends up exploiting minor weaknesses that no human would have payed attention to> False.> I was just saying that if the computer lets you get a strangly good position in the opening, it's not just playing stupid, in probably has something else in mind.

<The engines I looked at that involved self-learning would have all of the evaluation function set in place with only the values (the eventual 'constants')> The ones I've heard of do not have the peice values as constants (that are changing), but instead have a formula in place which is made up of the parts that give a peice its value (number of squares attacked, number of squares defending, mobility, escape options, etc.) and the computer "learns" how much each of these factors impacts each peice, and then finds the value of peices in each situation, so the peice values are never constants, but more accurate depending on the situation.

We're saying the same thing, though maybe I didn't make it obvious enough. However, note that with what you are saying now, the computer is basing it's view of the value of a piece only on certain predetermined factors. It cannot make up what factors it wants to consider, so it would see a bishop executing a pin and be happy. How is it able to guess how useful that will be in the long run?

<<So you're saying that you can find better openings by assuming your opponent will play worse moves?> No, I'm not saying that at all. What I'm saying is that I can safely assume that no human that I play will ever find a move that gets an advantage in every line after 15 perfect moves, which a very strong computer might (but probably won't) see. Instead, humans will only play according to what they see in the near future, or advice they have learned (like reading in an opening book that a certain move will give an endgame advantage).>

So what's the vaule in your statement? Are you actually considering playing less than the best move at some point hoping your opponent will make an even worse move back? How can you plan ahead and make moves without assuming your opponent will play back perfectly?

<the computer is basing it's view of the value of a piece only on certain predetermined factors.> Yes, but the how much each factor impacts the relative value of a peice is yet to be determined. Through practice and learning the computer will weight each factor according to the impact it has learned it has, and then apply the weighted system to the peice, so the factors aren't really predetermined; the computer just knows what they are.

<The only way we make computers decent at chess is really by hardwiring these long term concepts into it.> True, but as I have said with AI engines, you can still do this, but instead of giving each "concept" a relative value, leave that up to the computer to learn with practice, which can lead to good evaluations of positions.

<I have no doubt that if some human wanted to they could specialize in anti-computer play> Absolutely! I am actually very good at this myself, because I play against a portable computer engine I have at home 3-4 times a day, and I have done a lot of research on anti-computer play. Like you have mentioned, at this point in time turning off the opening book in chess engines would be a disadvantage, so I have learned a lot of rare variations in an attempt to get it out of the opening book. However, you have also said that you could probably beat a chess computer in the opening if its book was off, but this is not the reason I do this. I get out of the opening book to make the computer waste time thinking in the early opening, so that its play in the middle game is not as strong because it is out of time to think. This is the main reason that computers are weak with their opening books turned off: not because they play bad (although they might) but because they run out of time thinking.

<it is not uncommon for the computer to evaluate a human move as bad and then after a Computer reply without blundering the human makes the next move and computer evaluates the position as very bad(blunder) and then after the human makes their reply the Computer evaluates the game as won for the human!>

This is very common. However, the situation you refer to sounds like one of those sacrifices where the computer can't see far enough ahead to see the compensation that will come of it. In this case, the solution is just to give the computer a little more time. Example: once I was playing the Exchange Ruy Lopez against a relatively weak computer under a time limit, and we played down the main line to the point where I pinned his king's knight, he played h3, and I played h5. If you know this line then you will know that for white to capture the bishop is a devestating trap that will lead to checkmate in just a few moves. However, the computer could not see this far ahead and took the bishop, giving itself a won position. Two moves later, it gave me a won position and I mated it the next move. My point is that this was not a problem with the computer, it was a problem with the time. If you have read through the last few pages of this debate then you will know that we were talking particularily about computers that evaluate, NOT computers that play. In that case, there is no time limit, and these short traps will not occur.

<This can happen because the human has a particular strategy in a certain opening where he/she needs to post pieces on particular squares whilst allowing for a particular temporary weakness in his camp where the computer at first judges the moves to be bad but later in the resulting position evaluates that the human's position is indeed won.>

I like the point that you pointed out the Yugoslav Attack for an example. The Yugoslav is an opening that looks like just normal developing moves, but really has a plan for mate if black castles king-side. Another example is the English Attack against the Najdorf Sicilian, which can be deadly if black castles king-side. Humans no from experience to keep their king in the center of the board, but a computer that is not powerful enough under the time limit can sometimes make the mistake of castling and then a few moves later it gives itself a lost position. Now this is interesting because there are two solutions: the first is the obvious one. Give the computer more time. The second is available only for self-learning AI computers that we were talking about. After a few lost games of castling king-side in the English Attack, they will stop making that mistake. Situations like these are what make self-learning computers a hopefull prospect for the future.

<one cannot trust Computer evaluation 100% as how they analyse and how we humans play is different.>