by Steve Lopez

Oooooooooo, boy, I can already picture the e-mails that'll pour in on this one...

Once in a while (a couple of times in just the last few months, in fact) I find myself in a discussion of how chess programs evaluate opening gambits. It usually starts out with a user stating that "chess computers always evaluate gambits incorrectly. Don't you agree?" To which my answer is a clear, unequivocable, "Wellllllll, yes and no."

You have to cut me a little slack here, because there's no simple answer to the question. There are two key points that make the answer more than a bit ambiguous:

  1. The nature of gambits
  2. The algorithm of the program in question

I've written extensively on the subject of gambits (in fact, they're the main topic of my personal webpage and the subject of entire books written by others), so I won't try to rehash it all here. However, you can boil it down to a simple idea: in a gambit opening, you're giving up material (usually a pawn) in exchange for some non-material consideration, such as a development advantage, an attack, or some other positional consideration. In other words, you sac a pawn to get some other type of advantage. However, that advantage usually decreases over time, so you have to be quick in exploiting it.

So, almost by definition, a gambit by nature gives you an advantage but it's one that doesn't last forever. A great example is that old standby, the Danish Gambit: 1.e4 e5 2.d4 exd4 3.c3 dxc3 4.Bc4 cxb2 5.Bxb2

White has chucked two pawns overboard for a development advantage, control of the center, and an attack (both Bishops are pointed right at Black's vulnerable Kingside). But this advantage won't last through the whole game. Black's going to try to nullify the advantage by playing ...d5 and following it up by developing his own pieces aggressively. So who's ahead here? Right now, it's White. But Black's going to do his best to equalize. The ultimate verdict on this opening's theory depends on who you believe. Traditionalists say that Black can easily equalize and probably come out ahead. John Nunn (in Nunn's Chess Openings) gives an edge to White or, at worst, an unclear position (provided White knows his stuff).

But fire up a chessplaying program and have it analyze the position. Every program that I've had chew on this position says that White is stinking up the joint. If the position's OK for White, why do programs say otherwise?

When you get right to the heart of the matter, chessplaying programs are bean counters, older programs moreso than the newer ones. Older chess engines are primarily concerned with material -- which player has more wood on the board. So an older program looks at this position, see a two pawn deficit for White, and immediately says Black's winning big. Newer programs analyze a bit differently. Over the years, programmers figured out that the material count can often be misleading and have been building positional knowledge into their programs -- factoring in elements such as King safety, piece mobility, pawn structure, etc. However, the programs are still essentially bean counting: assigning numerical values to not just the material but to various positional factors as well.

This brings us to an interesting phenominon. Although all chess playing programs will see White as being behind in this position, the newer ones will show White as being behind by less than two pawns. Keep this in mind -- we'll come back to it in a minute.

Going back to our Danish Gambit position, we see that White has the advantages we discussed a few paragraphs back. But Black's not going to sit on his hands here; he's going to fight tooth and nail to nullify this advantage. If, for example, Black simply decides to adopt a "Moroccan rug merchant" strategy, trading off pieces and furiously hoovering the board so that all the pieces come off, he gets White into a losing endgame. If the pawn count remains the same, but the pieces come off, White is two pawns down in the ending and will certainly lose (unless his opponent is a rhesus monkey). That's the danger of gambit play -- you've give up material and if you can't win it back or exploit your advantage to get an 1800's-style quick mate, you're dead when the endgame rolls around. This is what I meant earlier by "the nature of gambits". In most gambits you get an early non-material advantage. But this advantage erodes over time -- little by little it slips away. You start off being one pawn down, but with better development or an attack. After awhile, if that development advantage doesn't pay off or your attack fails, you're just down a pawn -- period.

And this is why I contend that computers don't necessarily misevaluate gambits. It really depends on the particular engine and the phase of the game in which the position occurs.

Returning to our Danish Gambit position, I had a few chess engines chew on the position and I noticed something interesting. None of the newer engines I used (such as Fritz6, Hiarcs7.32, or Gambit Tiger 2.0) show this position as giving a full two pawn advantage to Black. The worst evaluation I saw was around 1.5 pawns and some were as "good" as a 1.20 advantage to Black. This is because of the positional criteria built into the engines. They're seeing that White does have some compensation for the material he gave up -- they just don't find it to be full compensation. And the evaluation varies from engine to engine: some engines are better than others at picking up on the non-material factors in a chess position. So are the engines "misevaluating" the position? To some extent, yes -- White has a few intangibles in exchange for the sacrificed pawns, intangibles that aren't being fully identified by the engines. To some extent, no -- White is down material and if he can't make his intangible advantages count (or if he can't at least recapture the pawns), he's going to be a hurting pup later in the game due to his material deficit.

In fact, "later in the game" is a crucial phrase here. As we've seen, the gambiteer's advantage erodes over time, so as the game progresses the engine's evaluation will tend to "normalize" -- that is, as the intangible (in a mathematical sense) advantage gets frittered away, the engine's evaluation becomes more accurate, until it's "spot on": the gambit player's position is a losing one because he's down material and he no longer has a non-material advantage. So this is what I mean when I refer to the "phase of he game". An engine might be misevaluating a position in the opening, but the evaluation becomes painfully (for the gambit player) accurate as the game moves into the middlegame and (especially) the endgame.

Further muddying the waters is the human's perception of gambits. Some players (like myself) love 'em, relishing the attacking possibilities and excitement they generate. Other players completely reject the idea, preferring instead to play in a safer, more conservative style in which they don't feel the need to risk the material (and a possible loss). In fact, you'll see frequent discussions/arguments/flamewars on the Interrant chess message boards over the idea of gambiteering. These varying opinions consequently color both sides of the engine evaluation argument. A player who thinks that the sacrifice of a pawn in the opening is never justified will tend to agree with the idea that chess engines evaluate such positions perfectly, while diehard gambit players take the position that engines frequently misevaluate gambit positions. Neither position is 100% correct; gambits per se are neither good nor bad -- that depends a great deal on the individual player's knowledge, style, and temperament. And a lot rides on the individual position.

Despite my personal love for gambits, I tend to be a fence-sitter in the debate. Saying that I enjoy gambit play is not the same thing as saying that I think all gambits are sound. Some (like the Fred: 1.e4 f5) are just plain bad, while many others (like the Danish, in my humble opinion) are perfectly playable. So in judging an engine's evaluation of a gambit position, we have to rely on our own chess knowledge to make the final determination rather than blindly following a playing program's numerical assessment (an idea that I've been harping on for the better part of fours years in ETN). When you make this judgement, keep in mind the "bean counting" nature of chess engines and keep an eye on the evaluation. If you give up two pawns in an opening but see an evaluation of "-1.25", it's a pretty fair bet that you have some level of compensation for the material. It's your job to figure out where that compensation lies and determine how best to exploit it, before that advantage slips away. And that's where the fun and excitement lies with gambits: in living on the edge and dicing with death.

So do chess engines misevaluate gambits? Sometimes yes, sometimes no. It's up to you to make the final determination on a case-by-case basis, based on your knowledge of a particular gambit and on chess engines in general. I hope this article has given you some information on the latter that will be useful in making these judgements.

As a side note (and, yes, I'll admit that I'm bragging here), my personal web page about gambits, the Chess Kamikaze Home Page, was mentioned by Al Lawrence in the July 2001 issue of Chess Life magazine as one of the "most interesting" on the Web. So I'd like to take this opportunity to publicly thank Mr. Lawrence for including it on his list, and I also want to thank all the folks who've offered comments, suggestions, questions, and ideas for improving the page.

Until next week, have fun!

You can e-mail me with your comments, suggestions, and analysis for Electronic T-Notes.