Think Like a Computer

Computers are very good at playing chess. But if I could examine a million positions per second, I’d be pretty good too. Throw in the ability to remember gigabytes of openings/endgames and I’m more than prepared for my next club championship… 🙂

There are many aspects of how a computer plays that we cannot or should not try to emulate - afterall, our brains work very differently. However, some aspects are interesting to compare and may even be instructive… it’s not just speed of “thought”…

Candidate moves

Possibly the most under-estimated skill in human chess, the skill of discovering candidate moves is critical to calculation ability. Humans are often prone to being drawn to too small a set of move options, and we don’t make enough effort to look for alternatives.

Alternatively, computers are good at looking at a wide range of moves; in fact, their weakness is usuallly that they look at too many and attempts to prune this wide range can be risky in terms of disgarding a good move. People are sometimes in awe at how deep a computer calculates but the breadth of the search is a key part too. Often it’s the computer’s short range tactical ability that is deadly – it simply considers a move that we don’t.

Iterative deepening

Further to candidate moves, computers are good at not calculating too deeply, too soon. So they search a range of moves at a depth of 8, before looking at any at 9, 10, etc. using “iterative deepening”. The deeper the search, the longer it usually takes. This means that they have to repeat some of their search, but this is vastly outweighed by avoiding spending too much time searching a given move when a better move exists. Spending a minute looking at 33.Qg4, before realising 33.Rb2 is checkmate, isn’t clever. This is an extreme example but humans will often start calculating too deeply before looking for earlier candidates. This may be ok in some cases, e.g. if the play is forced, but it can also turn out to be wasted effort if an earlier better move is discovered. Maybe we need more "iterative deepening".

Look at all checks/captures

Computers often put more emphasis on forcing moves such as captures and checks. In particular, they often don’t stop calculating a given line while there are further forcing moves. Disabling this approach can seriously weaken their play. I think this helps highlight the importance of paying attention to such forcing moves.

Registering change in position following a move

Humans need to visualise during calculation whereas computers are effectively making moves on a “physical board”. Ok, but more instructive is how following a move, a computer will “observe” the new position with total disregard for how it was arrived at - it won’t even factor in the last move! Us humans like to know what our opponent’s last move was and this helps to identify how the position has changed. However, we’re sometimes guilty of not fully registering all changes. Yes, we may see that the knight has gone to e5 but have you realised all the ranks/files/diagonals which have been opened/closed as a result? Sometimes we’d benefit by fully acknowledging all changes.

Playing for both sides

A computer shows no bias in terms of who it wants to win. It looks, with equal effort, for good moves for both sides. As a comparison, some players put more emphasis on finding ideas for themselves and it’s not always easy trying hard to refute a nice idea that you want to work. The benefits of being objective is clear.

Do not play blindly on general grounds

It’s all too easy to play similar positions in a similar manner. Sometimes having a slight change in the position may not matter. Or it may make a big difference. Again, computers are superior to humans in this respect since they always play on the basis of the specific position at hand. Of course, their brute force calculation is a key part in identifying whether a slight difference matters or not, but it’s still instructive to see how effective such specific play can be.


So, while we’re very different from computers, I don’t see why they should be ignored when considering what makes a strong player play well. Their approach to chess can add further recommendation to techniques that we have recognised as being useful in human terms. It’s not always that computers are simply faster; sometimes their “thought process” is simply better.

or in a nutshell, complete focus and unrelenting.

Hi V.

Interesting post.

"I don’t see why they should be ignored when considering
what makes a strong player play well. "

The strong chess player has something very important
no computer will ever have.

Without it they would not be strong players. It's essential.

Intuition.

Show me a move played by a computer based soley on intuition.

Yet practically all the famous (and 1000's not so famous) combinations
were based on intuition.

They never saw all the ramifications of a piece sac, but they just knew
it was sound and deadly.

No box will ever have that. It will have to analyse and analyse and may
dismiss it due to it's horizon.

You will surprised how little calculating good players actually do in their games.
Whereas a computing is number crunching all the time.

What can we learn from that?
We cannot analyse every postion like it does into dust.

You can add a sense of danger and judgement but I think intuition
covers everything a good chess players has that a computer
will never have.

And unless you are Capablanca or Tal you can only pick
this up by playing and playing and playing.

Originally posted by Varenka
Computers are very good at playing chess. But if I could examine a million positions per second, I’d be pretty good too

I never got beyond these opening sentences because I don't think the second is very likely.

Computers are not good at chess because they examine a million positions. They are good at chess because, for the most part, they evaluate positions correctly. The same applies for GMs. It's not that GMs see things less strong players do not (although sometimes they will), its because GMs exercise better judgement.

A good player is good, not for what is seen, but for what is selected

Almost all of this is actually practiced by good human players (i.e., advanced players), and the advice can be found in books written for human players, such as Dan Heisman's The Improving Chess Thinker: examine all checks, captures and threats: first, those your opponent has pending after his move (whether or not using the "null move" method, i.e., if my opponent could move immediately again), AND those you have available to you as your own move, AND those your opponent has after your tentatively selected move; when you find a good move, look for a better one; chess is concrete -- don't rely on handwaving general principles of play in an analytical (i.e., tactical) position; assume that your opponent WILL find the best move among those you examine for him; examine forcing lines in preference to others, and follow through sufficiently far (i.e., until the position becomes quiet and can be evaluated without the uncertainty of further forcing moves); don't examine a single line in great depth until you have examined that move and additional candidate moves at least superficially in order to determine if you have a mate or something that stands out as very promising, instead of spending all your time analyzing a single line in great depth only to find later that another move gives you a mate in two. (Some writers, Kotov among them, disagree with this method.)

Chess computers, to the extent that they conform to such behaviors, do so because these are recognized principles of good play and because such heuristics are programmed in. Of course, there are far too many moves even for a computer to examine in depth in any reasonable time, and so computers sometimes have a tactical horizon beyond which a terrible vulnerability exists that would be obvious to a human player looking at the board at that point, but which isn't even considered by the computer.

Also, programming computers to embody heuristics accurately under a wide variety of circumstances can be difficult. There was a game with Kasparov against Deep Blue where the whole room of grandmaster viewers groaned, again and again, because the computer failed to grasp a simple strategic principle and refused to make an obvious rook move. If a computer doesn't understand strategic principles, but simply analyzes possible moves (of which there may be many in an endgame where the board is mostly clear and a few mobile pieces can zip around), for tactical opportunities, by brute force examination of moves, it can easily miss moves that would be obvious to human players.

In fact, strategy is a major weakness in computer play which is why one of the strategems employed by human players is to move the game into closed and restricted positions with few tactical opportunities. Another strategem is making weird moves in the opening to take the computer out of book and force it to rely on its heuristics, which tend to be much weaker in opening play than the collective experience of countless grandmasters as encapsulated in an opening database.

Originally posted by atticus2
I never got beyond these opening sentences because I don't think the second is very likely.

Computers are not good at chess because they examine a million positions. They are good at chess because, for the most part, they evaluate positions correctly. The same applies for GMs. It's not that GMs see things less strong players do not (althoug ...[text shortened]... ise better judgement.

A good player is good, not for what is seen, but for what is selected

I fully agree but I think you're leaving out a very important element. Tactics!


You ever look at a tough tactical puzzle "White to move and win" for so long you finally give up? For many positions a player like me might be baffled (even knowing it's there and to look for it) but GM will just see it instinctively.

I doubt any player in the world has a snowballs chance against the top chess engines when it comes to tactics.

Let me put it this way... there are some things that...

i) computers are better at and we can't/shouldn't try to copy (e.g. calculate a million positions per second)

ii) computers are better at and we can/should consider copying (e.g. candidate moves)

iii) computers are worse at (e.g intuition, planning, etc.)


I was discussing (ii). And I know my list of principles is something that humans are already aware of. The point is whether our experience at trying to get a computer to play has added more recommendation to these techniques or not.

e.g. if programmers found it useful to tell a computer to pay extra attention to checks/captures then it adds even more backing to this principle. Or what do computers make of the material scale used by human beginners (pawn = 1, knight = 3, etc.)? Do they agree? It's this kind of feedback that I think can be instructive.

Hi Paralad,

Tactics is what they do best and they are do this very well.
Brilliant and unbeatable in a heavy tactical postion.

But it's all based on crunching the millions of positions it sees
and selecting the best line.

Humans cannot do that.
From an instructive point of view they are useless.

We use signposts, experinece and intuition to tell us when to stop
and look for the combination and do some calculating.

These things use brute force for every move. Every move.

Also they have no idea about the art of chess. They cannot entertain.
If they see a wonderful instructive Queen sac with mate in 4 or a cold easy
to see mate in 2 you will never ever see the Queen sac.

Step through this with the next button.


(Paralad never give up on any tactical problem you are trying to solve. If you
cannot see it then play what you would play in a game. Never give up.)

Hi V.

One thing in a modern computers favour is they are not blinded at all
by this Pawn=1 Knight = 3 Bishop = 3 naff nonsense.

Adding up the pieces off the board is a recipe for disaster and
stifles the imagination beyond repair.
It should not even be used as a guidline.

Knights are 'worth' more than Rooks in the opening, the e & d pawns
are 'worth' more than other pawns. etc etc.

The computer does not see these as units as numbers it sees them
as working values towards an end.
It will squeeze the best tactical use out of every piece and nowhere
in it's memory will a little counter be going up by one every time
it captures a pawn, or up by 5 when it spears a Rook.

It does what what we all should be doing and looking at what's
on the board and what position is infront of us.
Not counting points off the board.

V- Your post reminds me of Kotov. He wrote that Botvinnik was trying to teach a computer to play chess but instead we should try and think like a computer. This was the opening to the chapter on tree of variations.

Tisdall wrote very intelligently on why trying to think like a computer doesn't work well and that our fuzzy, intuition based approach to choosing candidate moves to calculate may in fact be better.

"I don't think like a tree, do you think like a tree?"

Hi Nimzo.

Have you got a source for the Tisdall article.
I believe you, I'd just like to see it.

Originally posted by greenpawn34

One thing in a modern computers favour is they are not blinded at all by this Pawn=1 Knight = 3 Bishop = 3 naff nonsense.

They can't calculate every line to mate, so how do they evaluate a position at the end of a line? They do factor in material values, albeit on a much more complex scheme than the 1,3,3,5,9 system. And sometimes they're too materialistic. e.g. Rybka suffered (maybe still does) from a wrong coloured rook pawn in bishop endings.

But that's going off the point. More relevant is the question: what if we experiment with different values in a more complex scheme (a bonus for the bishop pair; penalising doubled pawns; etc.) then how does the effect on the computer's play match our expectations? You see, it can be used to model certain guidelines or principles, and then the results compared. And the conclusion may be instructive for humans.

<duplicate post removed>

Originally posted by nimzo5
Tisdall wrote very intelligently on why trying to think like a computer doesn't work well

The title was indeed a play on Kotov's book, but aside from that, I don't suggest people try to think like a computer. In fact, in my initial post I acknowledge that our brains are very different from computers.

Instead, I was trying to highlight what thinking techniques in chess have been raised even higher, in terms of their benefit, by identifying that not only do the top players rely upon them, but that computers do too.

Originally posted by greenpawn34
Hi Paralad,

Tactics is what they do best and they are do this very well...unbeatable in a heavy tactical postion.

. . .

It will squeeze the best tactical use out of every piece and nowhere
in it's memory will a little counter be going up by one every time
it captures a pawn, or up by 5 when it spears a Rook.

It does what what we all should ...[text shortened]... t what's
on the board and what position is infront of us.
Not counting points off the board.

Well, not every chess computer or chess engine does tactics well. The strength of chess engines varies considerably. Of course, you're talking about strong chess engines, but I think you're exaggerating a bit when you say unbeatable in a heavy tactical position. Even chess engines can't examine all lines and don't try to, because they are working under time constraints; heuristics determine their move search parameters, and time constraints give them a tactical horizon beyond which number of moves in a given line they do not search. These do not always produce perfect results even in heavily tactical positions.

Someone else pointed out that they can overrate material gains at the expense of positional or other considerations, so to that extent they ARE counting points.

I found your thread-opener to be well-written and stimulating, Varenka. My point was, why not cut out the middle-man? You're really advocating the principles of master players, not that anyone think like a computer. Computers' use of them (as imperfectly embodied in programmable algorithms) is derivative; and asking human players to "think like a computer" may create a condition of psychological alienation that defeats or hinders their adoption. That said, your spin on the issue was entertaining and provocative.