It is more than a decade now that no human has beaten a top chess program, and now there are computer vs computer matches continuously running and getting results. Those results are used by programmers to improve the algorithms. The improvement is not merely a result of the increasing CPU speed, but also the result of theory research and improvement in the algorithms that prune and evaluate the game tree. This process was driven in the 80s and partly in the 90s by the desire to make the programs stronger against humans, but beyond a certain point the programs became so strong that now they are only tested against each other (and even against themselves) and I wonder if that is causing a deviation from chess into... another thing.

Hence my question. When top chess programs fight against each other at a level that no human could withstand, evaluating branches of the game tree up to an insane depth, how do we know that, beyond a certain depth, the moves make sense?

Let me explain my question with a simplified example. I bought recently a table chess computer in a toy shop chain, and it plays really bad. Even set at 3 minutes per move, I honestly doubt it is reaching farther than 1 or 2 plies at most and I am sure it is without any check extensions. It is really easy to beat, it goes out of the opening book at the 2nd or 3rd move, and it blunders miserably all the time. However, if you force it to play against itself, it stays within the opening book quite long, you don't see obvious blunders and the resulting game does not look bad. The algorithm playing against the same algorithm produces something that seems like decent chess, but I know it is fake, because if I interrupt it at any point and play against the machine I will beat it in no time. Hence my doubt when this example is translated to a much higher level... Do the strongest human see any value on those Stockfish-Houdini matches, do they give credibility to what those machines play, or it is simply crap nobody understands?

4 Answers 4


Do the strongest human see any value on those Stockfish-Houdini matches, do they give credibility to what those machines play, or it is simply crap nobody understands?

In many positions, computers play strange moves that no human would ever consider. People refer to moves like that as "computer moves." Some insanely deep tactic to reach a risky position that scores every so slightly higher than an obvious, safe move.

But, it's not crap. These computers playing these moves beat all human grandmasters. That's how we know the moves make sense, the computers would lose if their moves didn't make sense.

It's just that our human way of understanding the game is ultimately limited. Some moves just look like they go against all our understanding of positional play -- but they work. So it's our understanding that is wrong.

  • Yes, but we can only be certain that those computer moves are good enough to beat a human grandmaster. But beyond that, how do we know that they are not dancing around moves that are "only" enough to beat humans but nothing else. Let me show you a thought experiment where we use as "human" a computer limited to ply 3. You can have two other computers fighting in a match where they play moves that score very high within 3 plies, so that the "human" judges it as good chess, but those moves can be bad moves that, for instance, compromised the Queen at the 4th play. But, because...
    – Mephisto
    Commented Feb 4, 2017 at 22:14
  • But, because both computers do not attack each other and take advantage of that 4th ply, the game goes on and it makes always sense within a horizon depth of 3 plies. They can spend the whole match choosing moves that are bad when examined with a depth of 4 plies, but that seem good within 3 plies.
    – Mephisto
    Commented Feb 4, 2017 at 22:23
  • But there are many different chess engines that are better than grandmasters, and clear differences in strength between them. There are computers that play good enough to beat grandmasters, that are beaten by Fritz, which is crushed by Stockfish. And so on. And "only good enough to beat all humans but nothing else" is already pretty strong. Commented Feb 4, 2017 at 22:23
  • yes, it is pretty strong in absolute terms, but the increases in strength beyond human ability may be completely fake. It might be only a dance of moves that are "only" good enough to beat humans but with no real improvement.
    – Mephisto
    Commented Feb 4, 2017 at 22:25
  • 2
    Well if Stockfish beats all other computers, how is not a real improvement over weaker computers? There is no other objective, external meaning of what "real improvement" is. Commented Feb 4, 2017 at 22:29

Computers suffer from a fundamental flaw. (Or maybe it's humans with the problem . . . )

Let's suppose that we have a "perfect" computer, trained against similar "perfect" computers. On move one, it analyzes perfectly and realizes that the position is a draw. Moreover, it is so perfect that it realizes that with perfect play, 1.a4 leads to a draw, just as 1.e4 does. So, as White, it picks randomly and perhaps plays a4. As the game continues, it will continue playing "equal" moves until just before it reaches the point where one more such blunder would lose against perfect play. At that point it will start playing "perfect" moves to secure the draw.

Today's engines avoid this through most of the game because they can analyze only so deep. At some point they have to rely on position-evaluation heuristics, which usually give positions after 1.a4 a low score. But when we get to the endgame, at a certain point the analysis becomes "perfect" (thanks to exhaustive databases available of positions with a limited number of pieces) and this phenomenon kicks in. If we ever have a perfect computer, the process would begin at the very start.

A human eschews 1.a4 because he can never assume that he or his opponent is "perfect". A human is much more dependent on his ability to evaluate positions and therefore always tries to play "beautiful" chess.

Thus we reach the ironic conclusion. The more "perfect" our computers, the shoddier their apparent play. We have already reached the point where computers are of limited value in analysis. The engines routinely play seemingly ridiculous moves because they see threats that no human could calculate deeply enough to play over the board, while missing out on attacks because there happens to be a defensive line that is similarly impossible for humans to find. This isn't due to any lack of imagination, insight or effort by the programmers. On the contrary - it is because we have evolved our concept of "good" or "beautiful" chess to compensate for the limitations of our calculating ability.

  • If you give grandmaster a tablebase, he will improve his results against any human while he will be in the best possible shape against super opponent. The problem is not in the power of the engine but in bug inisde it nobody cares to remove.
    – hoacin
    Commented Feb 4, 2017 at 22:04
  • What an interesting explanation! (+1)
    – Mephisto
    Commented Feb 4, 2017 at 22:18

Yes, I believe there have been tournaments with highly-rated AI chess engines and after 300,000 games or so, some of the most incredible games were found with some incredible moves. But it's robust evaluation of the position through scoring and yes it is smarter than a human in this domain. But its world is limited into a 64 square area with billions of moves. The human brain is more adapted to the challenges and concepts that humans face every day which is different than a pretty massive domain like chess.

The AI in a sense is specialized in chess while humans are adapted to generic analysis to address many different challenges. Humans also specialize which is why we can sometimes see someone who is gifted in one area but terrible in another.

The world's best chess players do understand most computer moves but some computer moves are so deep that it's almost not even worth fully analyzing because it is so specialized move to that specific position while human players are more adapted to patterns and generic responses to a variety of positional strategic moves.

  • 1
    Great! Could you share a link to some of those amazing games?
    – Mephisto
    Commented Feb 20, 2017 at 19:27

Computers have consistently found strange defensive moves that have altered evaluations of many classic positions. Correspondence players still beat computers. The combination of human and computer is far more powerful than either alone.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.