Top engines without tablebases don't play as good as top GMs

[lech]

It's good you will evaluate it. There's always something to learn. The Endgame is very much a fertile ground of research for GM/s and for programmers of GM engines.

Lesson 1:
Stockfish needs 14 depths to breake a bonus for passed pawns in this simple position.

[d]8/8/8/8/4k3/7p/3PK2P/8 w - - 0 1

Code: Select all

Stockfish-222-32-ja:
  1/2	00:01	          16	12	+0,88	d2d3+ Ke4d4
  2/4	00:01	          67	51	+0,76	d2d3+ Ke4d4 Ke2d2
  3/4	00:01	         133	101	+0,88	d2d3+ Ke4d4 Ke2d2 Kd4d5
  4/6	00:01	         351	267	+0,92	d2d3+ Ke4d4 Ke2d2 Kd4d5 Kd2e3
  5/7	00:01	         875	658	+1,13	d2d3+ Ke4d4 Ke2d2 Kd4d5 Kd2e3 Kd5e5 d3d4+ Ke5d5
  6/8	00:01	       1.352	1.018	+1,13	d2d3+ Ke4d4 Ke2d2 Kd4d5 Kd2e3 Kd5e5 d3d4+ Ke5d5
  7/10	00:01	       2.962	2.230	+1,41	d2d3+ Ke4d4 Ke2d2 Kd4d5 Kd2e3 Kd5e5 d3d4+ Ke5d5 Ke3d3 Kd5e6
  8/13	00:01	       4.623	3.481	+1,57	Ke2f2 Ke4f4 d2d3 Kf4g4 Kf2e3 Kg4f5 d3d4 Kf5e6 Ke3e4 Ke6d6 Ke4e3
  9/14	00:01	       7.981	5.938	+1,77	Ke2f2 Ke4f4 d2d3 Kf4g4 Kf2e3 Kg4f5 d3d4 Kf5e6 Ke3e4 Ke6d7 Ke4e5 Kd7e7 Ke5e4 Ke7e6 Ke4e3 Ke6d6 Ke3e4 Kd6e6
 10/14	00:01	      13.053	9.712	+2,18	Ke2f2 Ke4f4 d2d3 Kf4g4 Kf2e3 Kg4f5 d3d4 Kf5e6 Ke3e4 Ke6d7 Ke4e5 Kd7e7 d4d5 Ke7d7
 11/14	00:01	      14.445	10.621	+2,18	Ke2f2 Ke4f4 d2d3 Kf4g4 Kf2e3 Kg4f5 d3d4 Kf5e6 Ke3e4 Ke6d7 Ke4e5 Kd7e7 d4d5 Ke7d7
 12/19	00:01	      29.183	21.224	+2,82	Ke2f2 Ke4f4 d2d3 Kf4g4 Kf2e3 Kg4f5 d3d4 Kf5e6 Ke3e4 Ke6d7 Ke4e5 Kd7e7 d4d5 Ke7d7 d5d6 Kd7d8 Ke5e6 Kd8e8 Ke6e5
 13/19	00:01	      33.683	24.496	+2,82	Ke2f2 Ke4f4 d2d3 Kf4g4 Kf2e3 Kg4f5 d3d4 Kf5e6 Ke3e4 Ke6d7 Ke4e5 Kd7e7 d4d5 Ke7d7 d5d6 Kd7d8 Ke5e6 Kd8e8 Ke6e5
 14/23	00:01	      61.564	43.786	+3,27	Ke2f2 Ke4f4 d2d3 Kf4g4 Kf2e3 Kg4f5 Ke3f3 Kf5e6 Kf3g4 Ke6e5 Kg4xh3 Ke5f4 d3d4 Kf4e4 Kh3g3 Ke4f5 h2h4 Kf5f6 h4h5 Kf6g5 Kg3f3
 15/28	00:01	     111.337	76.625	+5,05	Ke2f2 Ke4f4 d2d3 Kf4g4 Kf2e3 Kg4f5 Ke3f3 Kf5e6 Kf3g4 Ke6e5 Kg4xh3 Ke5f4 d3d4 Kf4e4 Kh3g3 Ke4f5 h2h4 Kf5f6 h4h5 Kf6g7 d4d5 Kg7f7 h5h6 Kf7g6 d5d6

In common sense it is better to have a higher passed pawn. But not in lots of pawn endgames.
It is good, if it starts from root. Imagine, problems on much higher ply.

Lesson 2:
It is possible to get this possition on a very high ply. It needs 20 additional depths to verify that it is a wrong choise.

[d] 8/8/4k3/8/3PK3/7p/7P/8 w - - 0 1

Code: Select all

Stockfish-222-32-ja:
  1/2	00:00	          14	22	+1,53	d4d5+ Ke6d6
  2/3	00:00	          50	80	+1,41	d4d5+ Ke6d6 Ke4d4
  3/4	00:00	         134	214	+2,06	d4d5+ Ke6d6 Ke4d4 Kd6d7
  4/6	00:00	         377	603	+1,97	d4d5+ Ke6d7 Ke4e5 Kd7e7 Ke5e4
  5/6	00:00	         682	1.091	+1,93	d4d5+ Ke6d7 Ke4d4 Kd7d6 Kd4e4 Kd6c5
  6/9	00:00	       1.916	2.989	+2,86	d4d5+ Ke6d6 Ke4d4 Kd6e7 Kd4e5 Ke7d7 d5d6 Kd7c6 Ke5e6
  7/10	00:00	       2.325	3.627	+2,90	d4d5+ Ke6d6 Ke4d4 Kd6d7 Kd4e5 Kd7e7 d5d6+ Ke7d7 Ke5d5 Kd7d8 Kd5e5 Kd8d7
  8/12	00:00	       2.865	4.367	+2,94	d4d5+ Ke6d6 Ke4d4 Kd6d7 Kd4e5 Kd7e7 d5d6+ Ke7d7 Ke5d5 Kd7d8 Kd5e6 Kd8e8
  9/13	00:00	       3.634	5.539	+2,86	d4d5+ Ke6d6 Ke4d4 Kd6d7 Kd4e5 Kd7e7 d5d6+ Ke7d7 Ke5d5 Kd7d8 Kd5e6 Kd8e8 Ke6e5 Ke8d7
 10/16	00:00	       6.117	9.324	+2,86	d4d5+ Ke6d6 Ke4d4 Kd6d7 Kd4e5 Kd7e7 d5d6+ Ke7d7 Ke5d5 Kd7d8 Kd5e4 Kd8e8 Ke4d4 Ke8d7 Kd4e5 Kd7c6 Ke5e6
 11/16	00:00	       6.558	9.758	+2,86	d4d5+ Ke6d6 Ke4d4 Kd6d7 Kd4e5 Kd7e7 d5d6+ Ke7d7 Ke5d5 Kd7d8 Kd5e4 Kd8e8 Ke4d4 Ke8d7 Kd4e5 Kd7c6 Ke5e6
 12/16	00:00	       8.601	12.799	+2,86	d4d5+ Ke6d6 Ke4d4 Kd6d7 Kd4e5 Kd7e7 d5d6+ Ke7d7 Ke5d5 Kd7d8 Kd5e4 Kd8e8 Ke4d4 Ke8d7 Kd4e5 Kd7d8 Ke5d5 Kd8d7 Kd5c5 Kd7d8 Kc5d4 Kd8e8 Kd4e4 Ke8d7 Ke4e5
 13/16	00:00	      10.599	15.772	+2,86	d4d5+ Ke6d6 Ke4d4 Kd6d7 Kd4e5 Kd7e7 d5d6+ Ke7d7 Ke5d5 Kd7d8 Kd5e4 Kd8e8 Ke4d5 Ke8d8
 14/20	00:00	      16.496	24.547	+2,86	d4d5+ Ke6d6 Ke4d4 Kd6d7 Kd4e5 Kd7e7 d5d6+ Ke7d7 Ke5d5 Kd7d8 Kd5c6 Kd8c8 Kc6d5 Kc8d7 Kd5c5 Kd7d8 Kc5c6
 15/22	00:00	      22.711	33.010	+2,74	d4d5+ Ke6d6 Ke4d4 Kd6d7 Kd4e5 Kd7e7 d5d6+ Ke7d7 Ke5d5 Kd7d8 Kd5c6 Kd8c8 Kc6b6 Kc8d8 Kb6b5 Kd8d7 Kb5c5 Kd7d8 Kc5d4 Kd8e8 Kd4c4 Ke8d8 Kc4b5
 16/22	00:00	      28.888	41.988	+2,74	d4d5+ Ke6d6 Ke4d4 Kd6d7 Kd4e5 Kd7e7 d5d6+ Ke7d7 Ke5d5 Kd7d8 Kd5e6 Kd8e8 Ke6d5 Ke8d7 Kd5c5 Kd7d8 Kc5c6 Kd8c8 Kc6b6 Kc8d8 Kb6b5 Kd8d7 Kb5c5
 17/24	00:00	      37.824	53.803	+2,74	d4d5+ Ke6d6 Ke4d4 Kd6d7 Kd4e5 Kd7e7 d5d6+ Ke7d7 Ke5d5 Kd7d8 Kd5c6 Kd8c8 Kc6b5 Kc8d7 Kb5c5 Kd7d8 Kc5c6
 18/24	00:00	      44.169	62.829	+2,74	d4d5+ Ke6d6 Ke4d4 Kd6d7 Kd4e5 Kd7e7 d5d6+ Ke7d7 Ke5d5 Kd7d8 Kd5c6 Kd8c8 Kc6b5 Kc8d7 Kb5c5 Kd7d8 Kc5c6
 19/24	00:00	      56.893	79.127	+2,74	d4d5+ Ke6d6 Ke4d4 Kd6d7 Kd4e5 Kd7e7 d5d6+ Ke7d7 Ke5d5 Kd7d8 Kd5c6 Kd8c8 Kc6b5 Kc8d7 Kb5c5 Kd7d8 Kc5d4 Kd8e8 Kd4c4 Ke8d8 Kc4d5 Kd8d7 Kd5e5 Kd7d8 Ke5d4
 20/28	00:00	     183.986	226.305	+1,69	d4d5+ Ke6d6 Ke4d4 Kd6d7 Kd4e5 Kd7e7 Ke5f5 Ke7d7 Kf5e5
 21/28	00:00	     417.863	445.482	 0,00	d4d5+ Ke6d6 Ke4d4 Kd6d7 Kd4e5 Kd7e7 d5d6+ Ke7d7 Ke5d5 Kd7d8 Kd5e4 Kd8d7 Ke4e5 Kd7d8 Ke5d5 Kd8d7 Kd5c5 Kd7d8 Kc5c4 Kd8e8 Kc4d4 Ke8d8 Kd4d5

Both are very easy examples only.

[Don]

It's good you will evaluate it. There's always something to learn. The Endgame is very much a fertile ground of research for GM/s and for programmers of GM engines.

Lesson 1:
Stockfish needs 14 depths to breake a bonus for passed pawns in this simple position.

What is this supposed to prove? This position is a win for white and Komodo see's the answer all the way to checkmate in about 30 seconds.

[d]8/8/8/8/4k3/7p/3PK2P/8 w - - 0 1

Code: Select all

Stockfish-222-32-ja:
  1/2	00:01	          16	12	+0,88	d2d3+ Ke4d4
  2/4	00:01	          67	51	+0,76	d2d3+ Ke4d4 Ke2d2
  3/4	00:01	         133	101	+0,88	d2d3+ Ke4d4 Ke2d2 Kd4d5
  4/6	00:01	         351	267	+0,92	d2d3+ Ke4d4 Ke2d2 Kd4d5 Kd2e3
  5/7	00:01	         875	658	+1,13	d2d3+ Ke4d4 Ke2d2 Kd4d5 Kd2e3 Kd5e5 d3d4+ Ke5d5
  6/8	00:01	       1.352	1.018	+1,13	d2d3+ Ke4d4 Ke2d2 Kd4d5 Kd2e3 Kd5e5 d3d4+ Ke5d5
  7/10	00:01	       2.962	2.230	+1,41	d2d3+ Ke4d4 Ke2d2 Kd4d5 Kd2e3 Kd5e5 d3d4+ Ke5d5 Ke3d3 Kd5e6
  8/13	00:01	       4.623	3.481	+1,57	Ke2f2 Ke4f4 d2d3 Kf4g4 Kf2e3 Kg4f5 d3d4 Kf5e6 Ke3e4 Ke6d6 Ke4e3
  9/14	00:01	       7.981	5.938	+1,77	Ke2f2 Ke4f4 d2d3 Kf4g4 Kf2e3 Kg4f5 d3d4 Kf5e6 Ke3e4 Ke6d7 Ke4e5 Kd7e7 Ke5e4 Ke7e6 Ke4e3 Ke6d6 Ke3e4 Kd6e6
 10/14	00:01	      13.053	9.712	+2,18	Ke2f2 Ke4f4 d2d3 Kf4g4 Kf2e3 Kg4f5 d3d4 Kf5e6 Ke3e4 Ke6d7 Ke4e5 Kd7e7 d4d5 Ke7d7
 11/14	00:01	      14.445	10.621	+2,18	Ke2f2 Ke4f4 d2d3 Kf4g4 Kf2e3 Kg4f5 d3d4 Kf5e6 Ke3e4 Ke6d7 Ke4e5 Kd7e7 d4d5 Ke7d7
 12/19	00:01	      29.183	21.224	+2,82	Ke2f2 Ke4f4 d2d3 Kf4g4 Kf2e3 Kg4f5 d3d4 Kf5e6 Ke3e4 Ke6d7 Ke4e5 Kd7e7 d4d5 Ke7d7 d5d6 Kd7d8 Ke5e6 Kd8e8 Ke6e5
 13/19	00:01	      33.683	24.496	+2,82	Ke2f2 Ke4f4 d2d3 Kf4g4 Kf2e3 Kg4f5 d3d4 Kf5e6 Ke3e4 Ke6d7 Ke4e5 Kd7e7 d4d5 Ke7d7 d5d6 Kd7d8 Ke5e6 Kd8e8 Ke6e5
 14/23	00:01	      61.564	43.786	+3,27	Ke2f2 Ke4f4 d2d3 Kf4g4 Kf2e3 Kg4f5 Ke3f3 Kf5e6 Kf3g4 Ke6e5 Kg4xh3 Ke5f4 d3d4 Kf4e4 Kh3g3 Ke4f5 h2h4 Kf5f6 h4h5 Kf6g5 Kg3f3
 15/28	00:01	     111.337	76.625	+5,05	Ke2f2 Ke4f4 d2d3 Kf4g4 Kf2e3 Kg4f5 Ke3f3 Kf5e6 Kf3g4 Ke6e5 Kg4xh3 Ke5f4 d3d4 Kf4e4 Kh3g3 Ke4f5 h2h4 Kf5f6 h4h5 Kf6g7 d4d5 Kg7f7 h5h6 Kf7g6 d5d6

In common sense it is better to have a higher passed pawn. But not in lots of pawn endgames.
It is good, if it starts from root. Imagine, problems on much higher ply.

Lesson 2:
It is possible to get this possition on a very high ply. It needs 20 additional depths to verify that it is a wrong choise.

[d] 8/8/4k3/8/3PK3/7p/7P/8 w - - 0 1

You're kidding right? This is a simple draw and there is no right or wrong choices here. Let me give you a lesson:

lesson 1: Rook pawn draws as long as defender can get on f8.
lesson 2: If rook pawns are missing position is a draw anyway. This plays like a simple KPvsK due to locked h pawns.

Komodo see's this as a draw, although it does take several ply, but it takes only 2 or 3 seconds.

Code: Select all

Stockfish-222-32-ja:
  1/2	00:00	          14	22	+1,53	d4d5+ Ke6d6
  2/3	00:00	          50	80	+1,41	d4d5+ Ke6d6 Ke4d4
  3/4	00:00	         134	214	+2,06	d4d5+ Ke6d6 Ke4d4 Kd6d7
  4/6	00:00	         377	603	+1,97	d4d5+ Ke6d7 Ke4e5 Kd7e7 Ke5e4
  5/6	00:00	         682	1.091	+1,93	d4d5+ Ke6d7 Ke4d4 Kd7d6 Kd4e4 Kd6c5
  6/9	00:00	       1.916	2.989	+2,86	d4d5+ Ke6d6 Ke4d4 Kd6e7 Kd4e5 Ke7d7 d5d6 Kd7c6 Ke5e6
  7/10	00:00	       2.325	3.627	+2,90	d4d5+ Ke6d6 Ke4d4 Kd6d7 Kd4e5 Kd7e7 d5d6+ Ke7d7 Ke5d5 Kd7d8 Kd5e5 Kd8d7
  8/12	00:00	       2.865	4.367	+2,94	d4d5+ Ke6d6 Ke4d4 Kd6d7 Kd4e5 Kd7e7 d5d6+ Ke7d7 Ke5d5 Kd7d8 Kd5e6 Kd8e8
  9/13	00:00	       3.634	5.539	+2,86	d4d5+ Ke6d6 Ke4d4 Kd6d7 Kd4e5 Kd7e7 d5d6+ Ke7d7 Ke5d5 Kd7d8 Kd5e6 Kd8e8 Ke6e5 Ke8d7
 10/16	00:00	       6.117	9.324	+2,86	d4d5+ Ke6d6 Ke4d4 Kd6d7 Kd4e5 Kd7e7 d5d6+ Ke7d7 Ke5d5 Kd7d8 Kd5e4 Kd8e8 Ke4d4 Ke8d7 Kd4e5 Kd7c6 Ke5e6
 11/16	00:00	       6.558	9.758	+2,86	d4d5+ Ke6d6 Ke4d4 Kd6d7 Kd4e5 Kd7e7 d5d6+ Ke7d7 Ke5d5 Kd7d8 Kd5e4 Kd8e8 Ke4d4 Ke8d7 Kd4e5 Kd7c6 Ke5e6
 12/16	00:00	       8.601	12.799	+2,86	d4d5+ Ke6d6 Ke4d4 Kd6d7 Kd4e5 Kd7e7 d5d6+ Ke7d7 Ke5d5 Kd7d8 Kd5e4 Kd8e8 Ke4d4 Ke8d7 Kd4e5 Kd7d8 Ke5d5 Kd8d7 Kd5c5 Kd7d8 Kc5d4 Kd8e8 Kd4e4 Ke8d7 Ke4e5
 13/16	00:00	      10.599	15.772	+2,86	d4d5+ Ke6d6 Ke4d4 Kd6d7 Kd4e5 Kd7e7 d5d6+ Ke7d7 Ke5d5 Kd7d8 Kd5e4 Kd8e8 Ke4d5 Ke8d8
 14/20	00:00	      16.496	24.547	+2,86	d4d5+ Ke6d6 Ke4d4 Kd6d7 Kd4e5 Kd7e7 d5d6+ Ke7d7 Ke5d5 Kd7d8 Kd5c6 Kd8c8 Kc6d5 Kc8d7 Kd5c5 Kd7d8 Kc5c6
 15/22	00:00	      22.711	33.010	+2,74	d4d5+ Ke6d6 Ke4d4 Kd6d7 Kd4e5 Kd7e7 d5d6+ Ke7d7 Ke5d5 Kd7d8 Kd5c6 Kd8c8 Kc6b6 Kc8d8 Kb6b5 Kd8d7 Kb5c5 Kd7d8 Kc5d4 Kd8e8 Kd4c4 Ke8d8 Kc4b5
 16/22	00:00	      28.888	41.988	+2,74	d4d5+ Ke6d6 Ke4d4 Kd6d7 Kd4e5 Kd7e7 d5d6+ Ke7d7 Ke5d5 Kd7d8 Kd5e6 Kd8e8 Ke6d5 Ke8d7 Kd5c5 Kd7d8 Kc5c6 Kd8c8 Kc6b6 Kc8d8 Kb6b5 Kd8d7 Kb5c5
 17/24	00:00	      37.824	53.803	+2,74	d4d5+ Ke6d6 Ke4d4 Kd6d7 Kd4e5 Kd7e7 d5d6+ Ke7d7 Ke5d5 Kd7d8 Kd5c6 Kd8c8 Kc6b5 Kc8d7 Kb5c5 Kd7d8 Kc5c6
 18/24	00:00	      44.169	62.829	+2,74	d4d5+ Ke6d6 Ke4d4 Kd6d7 Kd4e5 Kd7e7 d5d6+ Ke7d7 Ke5d5 Kd7d8 Kd5c6 Kd8c8 Kc6b5 Kc8d7 Kb5c5 Kd7d8 Kc5c6
 19/24	00:00	      56.893	79.127	+2,74	d4d5+ Ke6d6 Ke4d4 Kd6d7 Kd4e5 Kd7e7 d5d6+ Ke7d7 Ke5d5 Kd7d8 Kd5c6 Kd8c8 Kc6b5 Kc8d7 Kb5c5 Kd7d8 Kc5d4 Kd8e8 Kd4c4 Ke8d8 Kc4d5 Kd8d7 Kd5e5 Kd7d8 Ke5d4
 20/28	00:00	     183.986	226.305	+1,69	d4d5+ Ke6d6 Ke4d4 Kd6d7 Kd4e5 Kd7e7 Ke5f5 Ke7d7 Kf5e5
 21/28	00:00	     417.863	445.482	 0,00	d4d5+ Ke6d6 Ke4d4 Kd6d7 Kd4e5 Kd7e7 d5d6+ Ke7d7 Ke5d5 Kd7d8 Kd5e4 Kd8d7 Ke4e5 Kd7d8 Ke5d5 Kd8d7 Kd5c5 Kd7d8 Kc5c4 Kd8e8 Kc4d4 Ke8d8 Kd4d5

Both are very easy examples only.

[Don]

Lesson 2:
It is possible to get this possition on a very high ply. It needs 20 additional depths to verify that it is a wrong choise.

[d] 8/8/4k3/8/3PK3/7p/7P/8 w - - 0 1

Ok, sorry I misunderstood the point you were making which is that this position can be seen near leaf nodes of a deep search and of course it would be evaluated incorrectly by most programs.

But I think you miss my point too. You guys keep presenting special cases that demonstrate something that we have known for years, that a static evaluation function can be unreliable and that zugzwang can be a problem. This does not prove that humans are better at endings.

Actually, I think you have something seriously wrong here. You assume that humans do not have to do any "thinking" to understand this position but that only computers do. But that is not correct.

Imagine showing a strong human a bunch of random positions on flash cards where they have to identify as quickly as possible whether the position is a win or a draw (or loss.) When this position arises it takes any human a brief amount of time to process the position in his brain. He knows that rook pawn draws and that the king is not in front of the pawn, but humans are slow and it takes perhaps half a second just for the position to fully impact the brain and be recognized, then he has to reason on these 2 facts and "see the light" and this is going to take at least a couple of seconds or more. He may even take another second or two to convince himself that the king can indeed make it to the f8 square in time although I think a super Grandmaster will know this at a glance.

So I checked this on Komodo and it only took 2.6 seconds before the score stabilized to a draw. We might quibble about how long it would take a Grandmaster to see, recognize, reason and score a position like this but I'll bet 2 or 3 seconds is a lower bound. In other words I'm not convinced that even this example designed to make computers look bad really does what you intended. They don't just "know" it's a draw, they still have to reason on the position and take a little bit of time.

However this is just one example, there are many more positions that are more complex than this to a human, but much simpler for a computer, for example little tactics that take a few ply but are non trivial - in which case Komodo would see, process, recognize and score in a tiny fraction of a second.

Humans do well in position with simple well known rules like the one you just presented but they cannot compete with computers any longer because it's now getting to the point where most of the "tough" cases are handled by sheer CPU power faster than a human can even glance at a position and do trivial reasoning on it.

Just being able to find positions where computers plays the wrong move is not the same as "proving" that humans play better.

[pichy]

Lesson 2:
It is possible to get this possition on a very high ply. It needs 20 additional depths to verify that it is a wrong choise.

[d] 8/8/4k3/8/3PK3/7p/7P/8 w - - 0 1

Ok, sorry I misunderstood the point you were making which is that this position can be seen near leaf nodes of a deep search and of course it would be evaluated incorrectly by most programs.

But I think you miss my point too. You guys keep presenting special cases that demonstrate something that we have known for years, that a static evaluation function can be unreliable and that zugzwang can be a problem. This does not prove that humans are better at endings.

Actually, I think you have something seriously wrong here. You assume that humans do not have to do any "thinking" to understand this position but that only computers do. But that is not correct.

Imagine showing a strong human a bunch of random positions on flash cards where they have to identify as quickly as possible whether the position is a win or a draw (or loss.) When this position arises it takes any human a brief amount of time to process the position in his brain. He knows that rook pawn draws and that the king is not in front of the pawn, but humans are slow and it takes perhaps half a second just for the position to fully impact the brain and be recognized, then he has to reason on these 2 facts and "see the light" and this is going to take at least a couple of seconds or more. He may even take another second or two to convince himself that the king can indeed make it to the f8 square in time although I think a super Grandmaster will know this at a glance.

So I checked this on Komodo and it only took 2.6 seconds before the score stabilized to a draw. We might quibble about how long it would take a Grandmaster to see, recognize, reason and score a position like this but I'll bet 2 or 3 seconds is a lower bound. In other words I'm not convinced that even this example designed to make computers look bad really does what you intended. They don't just "know" it's a draw, they still have to reason on the position and take a little bit of time.

However this is just one example, there are many more positions that are more complex than this to a human, but much simpler for a computer, for example little tactics that take a few ply but are non trivial - in which case Komodo would see, process, recognize and score in a tiny fraction of a second.

Humans do well in position with simple well known rules like the one you just presented but they cannot compete with computers any longer because it's now getting to the point where most of the "tough" cases are handled by sheer CPU power faster than a human can even glance at a position and do trivial reasoning on it.

Just being able to find positions where computers plays the wrong move is not the same as "proving" that humans play better.

I decided to play Versus Komodo the same ending a few move earlier at move in 3 minutes without tablebase and here is the result:
[Event "Pichy vs Komodo Ending"]
[Date "2012.03.25"]
[Round "?"]
[Black "Komodo3-64"]
[Result "*"]
[BlackElo "3050"]
[WhiteElo "2200"]
[TimeControl "0+180"]
[SetUp "1"]
[FEN "2b5/2k3p1/p1p2pBp/P2p4/1P1P4/5PP1/7P/4K3 w - - 0 1"]
[WhiteType "human"]
[BlackType "program"]

1. g4 Kd6 2. Kf2 Ke7 3. Kg3 Kd7 4. Kh4 Kd8 5. Kh5 Ke7 6. Bd3 Kf7 7. Bf1 Bb7
8. Bd3 Bc8 9. h3 Ke7 10. Kg6 Kf8 11. Kh7 Kf7 12. Bg6+ Kf8 13. Kh8 Bd7 14.
f4 Be6 15. f5 Bd7 16. Bh5 c5 17. bxc5 Bc6 18. Bg6 Ba4 19. Bh7 h5 20. gxh5
Bd7 21. Bg6 Bc6 22. Kh7 Bb5 23. h6 gxh6 24. Kxh6 Bf1 25. h4 Bb5 26. Kh7 Bc6
27. h5 Ba4 {Black resigns} 28. h6 *

[D]5k2/7K/p4pBP/P1Pp1P2/b2P4/8/8/8 b - - 0 28

[Uri]

Even with tablebases computers still lose to the top GMs. Human players still produce better chess moves than even the most high-end chess engines.

Even Paul Morphy made better moves than currently the strongest commerical chess engine and Paul Morphy lived at a time way before the current personal computers were even invented.

[Uri]

Even with tablebases computers still lose to the top GMs. Human players still produce better chess moves than even the most high-end chess engines.

Even Paul Morphy made better moves than currently the strongest commerical chess engine and Paul Morphy lived at a time way before the current personal computers were even invented.

[Uri]

Even with tablebases computers still lose to the top GMs. Top human Grandmasters, rated 2600 FIDE and beyond, still come up with better chess moves than even the most high-end chess engines.

Even Paul Morphy made better moves than currently the strongest commerical chess engines running on an Intel hexa-core processor and Paul Morphy lived at a time way before the current modern personal computers were even invented.

[Terry McCracken]

Lesson 2:
It is possible to get this possition on a very high ply. It needs 20 additional depths to verify that it is a wrong choise.

[d] 8/8/4k3/8/3PK3/7p/7P/8 w - - 0 1

Ok, sorry I misunderstood the point you were making which is that this position can be seen near leaf nodes of a deep search and of course it would be evaluated incorrectly by most programs.

But I think you miss my point too. You guys keep presenting special cases that demonstrate something that we have known for years, that a static evaluation function can be unreliable and that zugzwang can be a problem. This does not prove that humans are better at endings.

Actually, I think you have something seriously wrong here. You assume that humans do not have to do any "thinking" to understand this position but that only computers do. But that is not correct.

Imagine showing a strong human a bunch of random positions on flash cards where they have to identify as quickly as possible whether the position is a win or a draw (or loss.) When this position arises it takes any human a brief amount of time to process the position in his brain. He knows that rook pawn draws and that the king is not in front of the pawn, but humans are slow and it takes perhaps half a second just for the position to fully impact the brain and be recognized, then he has to reason on these 2 facts and "see the light" and this is going to take at least a couple of seconds or more. He may even take another second or two to convince himself that the king can indeed make it to the f8 square in time although I think a super Grandmaster will know this at a glance.

So I checked this on Komodo and it only took 2.6 seconds before the score stabilized to a draw. We might quibble about how long it would take a Grandmaster to see, recognize, reason and score a position like this but I'll bet 2 or 3 seconds is a lower bound. In other words I'm not convinced that even this example designed to make computers look bad really does what you intended. They don't just "know" it's a draw, they still have to reason on the position and take a little bit of time.

However this is just one example, there are many more positions that are more complex than this to a human, but much simpler for a computer, for example little tactics that take a few ply but are non trivial - in which case Komodo would see, process, recognize and score in a tiny fraction of a second.

Humans do well in position with simple well known rules like the one you just presented but they cannot compete with computers any longer because it's now getting to the point where most of the "tough" cases are handled by sheer CPU power faster than a human can even glance at a position and do trivial reasoning on it.

Just being able to find positions where computers plays the wrong move is not the same as "proving" that humans play better.

You make a good argument but even I know this is a draw without any calculations. The two patterns are that obvious and seen instantly. The time is less or no more than a second. For a human that is about as instant as it gets.

Zugzwangs of course need to be remedied and some top players do read many many positions, even complex ones at a glance. Kasparov was shown an ending and looked confused as he thought the position was far too easy and when shown it modified he smiled as he had to think, a little.

Seeing him look at deep endings at a glance over breakfast, solving them virtually on the spot is quite unnerving to mere mortals!

A decade ago he was like a very fast supercomputer to the unaided eye.

[pichy]

Lesson 2:
It is possible to get this possition on a very high ply. It needs 20 additional depths to verify that it is a wrong choise.

[d] 8/8/4k3/8/3PK3/7p/7P/8 w - - 0 1

Ok, sorry I misunderstood the point you were making which is that this position can be seen near leaf nodes of a deep search and of course it would be evaluated incorrectly by most programs.

But I think you miss my point too. You guys keep presenting special cases that demonstrate something that we have known for years, that a static evaluation function can be unreliable and that zugzwang can be a problem. This does not prove that humans are better at endings.

Actually, I think you have something seriously wrong here. You assume that humans do not have to do any "thinking" to understand this position but that only computers do. But that is not correct.

Imagine showing a strong human a bunch of random positions on flash cards where they have to identify as quickly as possible whether the position is a win or a draw (or loss.) When this position arises it takes any human a brief amount of time to process the position in his brain. He knows that rook pawn draws and that the king is not in front of the pawn, but humans are slow and it takes perhaps half a second just for the position to fully impact the brain and be recognized, then he has to reason on these 2 facts and "see the light" and this is going to take at least a couple of seconds or more. He may even take another second or two to convince himself that the king can indeed make it to the f8 square in time although I think a super Grandmaster will know this at a glance.

So I checked this on Komodo and it only took 2.6 seconds before the score stabilized to a draw. We might quibble about how long it would take a Grandmaster to see, recognize, reason and score a position like this but I'll bet 2 or 3 seconds is a lower bound. In other words I'm not convinced that even this example designed to make computers look bad really does what you intended. They don't just "know" it's a draw, they still have to reason on the position and take a little bit of time.

However this is just one example, there are many more positions that are more complex than this to a human, but much simpler for a computer, for example little tactics that take a few ply but are non trivial - in which case Komodo would see, process, recognize and score in a tiny fraction of a second.

Humans do well in position with simple well known rules like the one you just presented but they cannot compete with computers any longer because it's now getting to the point where most of the "tough" cases are handled by sheer CPU power faster than a human can even glance at a position and do trivial reasoning on it.

Just being able to find positions where computers plays the wrong move is not the same as "proving" that humans play better.

You make a good argument but even I know this is a draw without any calculations. The two patterns are that obvious and seen instantly. The time is less or no more than a second. For a human that is about as instant as it gets.

Zugzwangs of course need to be remedied and some top players do read many many positions, even complex ones at a glance. Kasparov was shown an ending and looked confused as he thought the position was far too easy and when shown it modified he smiled as he had to think, a little.

Seeing him look at deep endings at a glance over breakfast, solving them virtually on the spot is quite unnerving to mere mortals!

A decade ago he was like a very fast supercomputer to the unaided eye.

========================================================================================================================

I checked this position in youtube.com and realized that even a master like myself or used to be 19 years ago rated 2200, and since then have read several books on endings, plus I took a quick look at this position and the explanation and decided to play versus Komodo at 3 minutes per move using my AMD Athlon (tm) Duo Core Processor 4850e 4.50 GHz

http://www.youtube.com/watch?v=5SgNPwe_7B0

This is the initial position:
[D]2b5/2k3p1/p1p2pBp/P2p4/1P1P4/5PP1/7P/4K3 w - - 0 1"

[Event "Pichy vs Komodo Ending"]
[Date "2012.03.25"]
[Round "?"]
[Black "Komodo3-64"]
[Result "*"]
[BlackElo "3050"]
[WhiteElo "2200"]
[TimeControl "0+180"]
[SetUp "1"]
[FEN "2b5/2k3p1/p1p2pBp/P2p4/1P1P4/5PP1/7P/4K3 w - - 0 1"]
[WhiteType "human"]
[BlackType "program"]

1. g4 Kd6 2. Kf2 Ke7 3. Kg3 Kd7 4. Kh4 Kd8 5. Kh5 Ke7 6. Bd3 Kf7 7. Bf1 Bb7
8. Bd3 Bc8 9. h3 Ke7 10. Kg6 Kf8 11. Kh7 Kf7 12. Bg6+ Kf8 13. Kh8 Bd7 14.
f4 Be6 15. f5 Bd7 16. Bh5 c5 17. bxc5 Bc6 18. Bg6 Ba4 19. Bh7 h5 20. gxh5
Bd7 21. Bg6 Bc6 22. Kh7 Bb5 23. h6 gxh6 24. Kxh6 Bf1 25. h4 Bb5 26. Kh7 Bc6
27. h5 Ba4 {Black resigns} 28. h6 *

I don’t understand why Komodo made the bad move 16.Bh5 c5? I was expecting 16… Bc8, probably Komodo realized that
I was going to sacrifice my bishop by playing the novelty move Be8 forcing the King to take or I would had taken pawn c6 and the pawn d5 in the next two moves. or I touched some keyboard without realizing anyway Komodo if Komodo played the expected move Bc8, then I would had played Be8 and after Black king take my bishop White king take Kxg7 etc...


PS: After the game I analyzed it again and Komodo3 played the expected move 16.Bh5 Bc8
[D]5k1K/3b2p1/p1p2p1p/P2p1P1B/1P1P2P1/7P/8/8 b - - 2 16

This is the final position when Komodo3-64 resigned
[D]5k2/7K/p4pBP/P1Pp1P2/b2P4/8/8/8 b - - 0 28

[Don]

Lesson 2:
It is possible to get this possition on a very high ply. It needs 20 additional depths to verify that it is a wrong choise.

[d] 8/8/4k3/8/3PK3/7p/7P/8 w - - 0 1

Ok, sorry I misunderstood the point you were making which is that this position can be seen near leaf nodes of a deep search and of course it would be evaluated incorrectly by most programs.

But I think you miss my point too. You guys keep presenting special cases that demonstrate something that we have known for years, that a static evaluation function can be unreliable and that zugzwang can be a problem. This does not prove that humans are better at endings.

Actually, I think you have something seriously wrong here. You assume that humans do not have to do any "thinking" to understand this position but that only computers do. But that is not correct.

Imagine showing a strong human a bunch of random positions on flash cards where they have to identify as quickly as possible whether the position is a win or a draw (or loss.) When this position arises it takes any human a brief amount of time to process the position in his brain. He knows that rook pawn draws and that the king is not in front of the pawn, but humans are slow and it takes perhaps half a second just for the position to fully impact the brain and be recognized, then he has to reason on these 2 facts and "see the light" and this is going to take at least a couple of seconds or more. He may even take another second or two to convince himself that the king can indeed make it to the f8 square in time although I think a super Grandmaster will know this at a glance.

So I checked this on Komodo and it only took 2.6 seconds before the score stabilized to a draw. We might quibble about how long it would take a Grandmaster to see, recognize, reason and score a position like this but I'll bet 2 or 3 seconds is a lower bound. In other words I'm not convinced that even this example designed to make computers look bad really does what you intended. They don't just "know" it's a draw, they still have to reason on the position and take a little bit of time.

However this is just one example, there are many more positions that are more complex than this to a human, but much simpler for a computer, for example little tactics that take a few ply but are non trivial - in which case Komodo would see, process, recognize and score in a tiny fraction of a second.

Humans do well in position with simple well known rules like the one you just presented but they cannot compete with computers any longer because it's now getting to the point where most of the "tough" cases are handled by sheer CPU power faster than a human can even glance at a position and do trivial reasoning on it.

Just being able to find positions where computers plays the wrong move is not the same as "proving" that humans play better.

You make a good argument but even I know this is a draw without any calculations. The two patterns are that obvious and seen instantly. The time is less or no more than a second. For a human that is about as instant as it gets.

Zugzwangs of course need to be remedied and some top players do read many many positions, even complex ones at a glance. Kasparov was shown an ending and looked confused as he thought the position was far too easy and when shown it modified he smiled as he had to think, a little.

Seeing him look at deep endings at a glance over breakfast, solving them virtually on the spot is quite unnerving to mere mortals!

A decade ago he was like a very fast supercomputer to the unaided eye.

========================================================================================================================

I checked this position in youtube.com and realized that even a master like myself or used to be 19 years ago rated 2200, and since then have read several books on endings, plus I took a quick look at this position and the explanation and decided to play versus Komodo at 3 minutes per move using my AMD Athlon (tm) Duo Core Processor 4850e 4.50 GHz

http://www.youtube.com/watch?v=5SgNPwe_7B0

This is the initial position:
[D]2b5/2k3p1/p1p2pBp/P2p4/1P1P4/5PP1/7P/4K3 w - - 0 1"

You must be joking. That is a terrible position for black, white has a clear plan here that I think must be winning. This is how you prove that humans are superior in the endgame?

I'll bet if you play komodo vs komodo white will win every time.

[Event "Pichy vs Komodo Ending"]
[Date "2012.03.25"]
[Round "?"]
[Black "Komodo3-64"]
[Result "*"]
[BlackElo "3050"]
[WhiteElo "2200"]
[TimeControl "0+180"]
[SetUp "1"]
[FEN "2b5/2k3p1/p1p2pBp/P2p4/1P1P4/5PP1/7P/4K3 w - - 0 1"]
[WhiteType "human"]
[BlackType "program"]

1. g4 Kd6 2. Kf2 Ke7 3. Kg3 Kd7 4. Kh4 Kd8 5. Kh5 Ke7 6. Bd3 Kf7 7. Bf1 Bb7
8. Bd3 Bc8 9. h3 Ke7 10. Kg6 Kf8 11. Kh7 Kf7 12. Bg6+ Kf8 13. Kh8 Bd7 14.
f4 Be6 15. f5 Bd7 16. Bh5 c5 17. bxc5 Bc6 18. Bg6 Ba4 19. Bh7 h5 20. gxh5
Bd7 21. Bg6 Bc6 22. Kh7 Bb5 23. h6 gxh6 24. Kxh6 Bf1 25. h4 Bb5 26. Kh7 Bc6
27. h5 Ba4 {Black resigns} 28. h6 *

I don’t understand why Komodo made the bad move 16.Bh5 c5? I was expecting 16… Bc8, probably Komodo realized that
I was going to sacrifice my bishop by playing the novelty move Be8 forcing the King to take or I would had taken pawn c6 and the pawn d5 in the next two moves. or I touched some keyboard without realizing anyway Komodo if Komodo played the expected move Bc8, then I would had played Be8 and after Black king take my bishop White king take Kxg7 etc...


PS: After the game I analyzed it again and Komodo3 played the expected move 16.Bh5 Bc8
[D]5k1K/3b2p1/p1p2p1p/P2p1P1B/1P1P2P1/7P/8/8 b - - 2 16

This is the final position when Komodo3-64 resigned
[D]5k2/7K/p4pBP/P1Pp1P2/b2P4/8/8/8 b - - 0 28