5/24/2020 Back in the 1980s some program authors tried to avoid the problems of the horizon effect by following suspicious lines a few half moves deeper. It became ever more difficult to use the horizon trap on them. The authors were not always forthcoming about their progress, so Frederic Friedel devised a number of test positions to find out more and to find how individual programs worked.

In my previous article I told you about the horizon effect, how in early days chess programs would often try to avert impending doom by pushing the threats over their calculation horizon. It used to be fun tricking chess computers by tempting them to giving up material in order to delay the inevitable. Here are some examples:

The first position is not really serious. I have tried to simulate the experiments of the 1980s by giving the diagram engine just a tenth of a second thinking time per move but even then it plays 1...Kf2, ...Ke3, ...e4 and wins. In 1983 a number of programs tried to prevent the white a-pawn from promoting by sacrificing pawns! 1.a4 f5? (1...Kf2! 2.a5 Ke3 3.a6 e4 and Black wins easily) 2.Bxf5 g6? 3.Bxg6 e4? 4.Bxe4 d3? 5.Bxd3 c2? 6. Bxc2 b1=Q+ 7.Bxb1 and now it is White that need eleven moves to mate. Quite hilarous.

The second position is one of my favourite studies by Leonid Kubbel. If you have not seen it before I urge you to stop reading now and take a shot at solving it yourself. It looks like Black will promote his a-pawn without any difficulty the bishop cannot stop it, since the al-h8 diagonal is blocked by the black d-pawn. The position in fact looks lost for White. But there is a clear win. Let us see: 1.Bf6 a2 2.c3 a1Q (of course not 2...dxc3 3.Bxc3+) and Black wins.

How about 1.Bf6 a2 2.c4+. That doesn't work either: 2...Kxc4 3.d3+ Kxd3 and the a-pawn promotes without problems. What if we play 1.c4+ right away? After 1...Kxc4 2.d3+ Kxd3? 3.Nc6 threatens a fork and allowing 4.Bf6 with a draw) Black will promote and win. But Black can play 2...Kd5! with smooth promotion.

Then how about knight moves? 1.Nd7 a2 2.Nb6+ Kc6/e6 wins quickly. 1.Nc6 looks interesting because of the fork threatened on b4, but what happens after 1...Kxc6 2.Bf6 Kd5? So things look bleak do you still believe White can win?

In 1983 I gave the Kubel study to one of the best table-top computers. It started with 1.Kb5 and soon switched to 1.c4. After an hour it began considering 1.Bf6. Half an hour later it became restless and kept switching between 1.Le7, 1.Kb5 and 1.Se7. Obviously it had realized the full seriousness of the situation. But then it suddenly switched to 1.Sc6. That was interesting it was considering sacrificing a knight. Had the computer actually found the very clever solution to the Kubbel study? 1.Nc6! Kxc6 2.Bf6 Kd5 3.d3!! a2 4.c4, Kc5 (the a-pawn will queen, but:) 5.Kb7! alQ and 6.Le7#!

I was deeply impressed: did the machine really see the mate threat at eleven ply? Computers at the time could not look that deep. However, one of my readers, Dr. Timm Deeg, proved, by moving the wK and wN around, that the computer hadnt an inkling why 1.Nc6 worked. It was the horizon effect! The computer was simply pushing the promotion (bad for White) beyond its search horizon. Letting it work on the Kubble study for a full night made the computer switch back to 1.Kb5 (losing).

So this time, coincidentally, the horizon effect helped the program, which played the right moves (after an hour and a half) for the wrong reasons. Today, any program will find 1.Nc6 in a second or two, then go down 30-40 ply and announce mate in 16-40 moves. My how times have changed!

So we have poked fun enough on ancient computers and programs. Let's turn our attention to later generations. In the new century (millennium!) we saw puzzles which could be used to fool certain programs, while others could find the solution with a little help.

Study number three, by Matous, is something you can experiment with. Until recently many programs were in general content with 1.Qd6+ Kg8 2.gxh7+ Kh8 3.Bxa5, which hardly yields more than a draw. There is, however, a very clear and forced win with a dramatic second move that is thrilling to see. Can you and your Fritz, Stockfish, or whatever, work it out?

"Send us an insoluble study," we told our readers, back in 2009. We wanted to know if there are positions which our silicon friends cannot solve. The puzzles must be elegant and the solutions readily comprehensible to humans. But they should cause a major headache (CPU-ache?) to computers.

The Hasek study was sent to us by Manuel Rodriguez from the Dominican Republic. "Some time ago I left Deep Rybka 3.0, dual core 2.0 GHZ with 3 GB RAM to analyse it for 36 consecutive hours," he wrote. "Rybka didn't find anything. I haven't tried the Monte Carlo function yet maybe that will produce the correct first move?!"

That was over ten years ago. Can our readers solve this puzzle with modern-day computers (100 times faster) and chess engines (brute force or AI)? It is still a genuine challenge for our electronic friends but for humans as well. The best way for the latter to handle this kind of position is to play the position out against a computer and find a way to hold a draw. It will tear you to pieces if you do not play precisely, play the correct and only strategy to hold a draw. The computer may not realise that it can never win the position, but you as a human will.

I look forward to comparing results. I will post them here in the coming week.

Here is the original post:

Darkness on the Horizon (2) - Chessbase News