AN EXPERIMENT

Neil Berch

The use of computers is an issue of great controversy in the correspondence chess community. Some believe that computers are the death of cc; others believe that they will enhance it. The strongest claims made are that computers will one day be able to defeat the best cc players. This claim, in my opinion, is ridiculous, because the best cc players could then aid their own play through selective computer use. A real question, however, is how strong computers by themselves are right now in cc play. Some have tried to answer this question in different ways. One Israeli player claimed that he had achieved tremendous success through the use of multiple computers, only intervening when the computers disagreed. This still required some judgment, however. People appear to be more interested in the question of just how far a so-called "postman" (someone with little or no ability who simple used the strongest computer available and the best software available) could go in top-level cc. One answer comes from a match between US player Steve Ham (ICCF 2498) and two of the better computer engines. That match, detailed on Franklin Campbell’s web page, ended in one win for the computers and three draws (a performance rating for the two computers of almost 2600). However, Steve handicapped himself by refusing to use any anti-computer strategies, and the number of games played was too small to really say anything significant.

I undertook a different sort of experiment, sort of by accident. Not thinking very well, I entered four e-mail sections of the ICCF World Cup last year. When the pairings for 40 games arrived, I contemplated what I had done. I had another 10-15 e-mail games going, and family and work responsibilities. I decided to only play two of those sections (leaving me with a still unmanageable 34 or so games). For the other two sections, I decided that I would conduct an experiment. I would be the postman for my version of Fritz 5. Needless to say, I told none of my opponents, as I wanted this to be as pure an experiment as possible. I randomly selected two sections (20 games) in which Fritz would make all of the decisions. Here are the ground rules I decided on:

1. In most games, I would play from Fritz’s opening book until it had either run out or reached the point where only losing games for my side were recorded.

2. In a few games, I substituted an opening book of games in Larsen’s Opening and the English Defense, both of which I was playing at the time (one might surmise that this would weaken the results, but it appeared to have very little effect).

3. After getting out of the opening, I would run Fritz for 15 minutes on each move (I needed my computer for work) and then play the best move. If the choice was indeed obvious, I might cut them time a bit. On occasion, I had an errand to run, and I might have been away for 30-40 minutes. Generally, however, I observed a 15 minute deadline. When I varied from that deadline, it had nothing to do with the chess position; it had to do with my life at the time.

4. I would resign a position if Fritz evaluated it at –2.5 pawns or more for two consecutive moves.

5. I would only offer draws in positions where I had repeated the position.

6. I would accept draw offers if Fritz evaluated my (our) position as inferior.

Well, the results are now in. In those two sections (65 and 67), I have completed play with a total of 8 wins, 4 losses, and 7 draws (11.5/19, plus one win by forfeit when the player was a silent withdrawal at the start). Given the strength of opposition, this equates to a performance rating of about 2285 (counting only the 16 games against opening who had ratings in the latest rating list, though counting the other three games has no effect). In one section, I will finish with 6.5/10 (including the forfeit win); in the other, 6/10. In both cases, this should be good enough for about a 3^rd or 4^th place finish. More interesting is the breakdown by opponent’s rating:

Opponent <2200: +3 –0 =3 (75%)

Opponent 2200-2300: +4 –1 =2 (71%)

Opponent 2300+: +1 –3 =2 (33%)

So, it appears that up to 2300, this version of Fritz (not the latest) on my hardware (Pentium Pro, with 32 mb of RAM, and at 233 Mhz—clearly not the latest) can certainly hold its own. Against high-level opponents, however, the results are much less impressive.

We can then try to extrapolate to answer the question of what sort of performance the ultimate postman could expect. In other words, what performance might we expect from someone with the best hardware, the best software, a longer period for the computer to calculate on each move, and perhaps better use of the Openings book?

Let me address each of these in turn:

1. The longer period for reflection is not likely (I believe) to have a significant effect. I ran a few of the closer decisions by Fritz for several hours, and the move choice did not change. No doubt it would occasionally improve, but I doubt that this would make a huge difference. My guess is that this would raise the performance by 25-50 rating points (which would have meant an additional half-point or perhaps two half-points out of the 19 games).

2 and 3. I consulted the latest Swedish Computer Chess rating list. Fritz 5 running on hardware similar to mine has a rating of 2459. The very best rating is obtained by Deep Fritz, running on much better hardware (128 mb RAM, 450 Mhz, with some other notations I don’t even understand). Its rating was 2653, a difference of 194 points. I might guess that the very best hardware could push this advantage to 225 points. However, this is in OTB conditions at 40 moves in 2 hours. I would expect that the difference would be much less in cc conditions. How much so is subject to conjecture, but I have a hard time imagining that Deep Fritz running on the best hardware would score worse than 60% against my Fritz 5 running on my hardware. I also have a hard time picturing Deep Fritz scoring better than 75%. If we round away from both of those figures a bit, we might suggest that the best software and hardware would be worth 90-170 points in cc.

4. Finally, eliminating the use of inferior openings might have had a small effect. Let’s offer up to 25 points for that (meaning Fritz would have picked up an extra half-point in one of the 8 games where I employed the English Defense or Larsen’s Opening).

So, if we add it all up, we might expect that a postman with all the latest software and hardware, as well as a good deal of computer time, would be able to exceed my performance by 115-245 points. If we treat my results as a range (2235-2335), we can then say that the current best postman’s range is 2350-2580. I think more realistically, it is not at either extreme of that range, so I would say that someone with the best software and hardware is likely to perform between 2400 and 2525. This would probably be good enough to win an ICCF Master section with some perseverance, but not good enough to win a World Championship semifinal. As computers get better, the postman will "play" better chess, but so will the top-level players who also use computers to blundercheck. So, at the very top, I think it is unlikely that someone can get within a couple of steps of the World Championship in cc without significant chess ability. Still, I was a little surprised at how Fritz 5 performed.

Specifically, Fritz found sacrifices with long-term effects, played endings somewhat better than expected, and showed more "creativity" than I would have thought. Here are a few games of interest:

Here is the only win against a player rated over 2300:

ICCF WC12/065

Berch - Wosch (2333)

1.b3 e5 2. Bb2 Nc6 3. e3 Nf6 4. Bb5 d6 5. Ne2 Bd7 6. O-O a6 7. Bxc6 Bxc6 8. d4 Qe7 9. c4 g6 10. Nbc3 Bg7 11. Qc2 O-O 12. dxe5 dxe5 13. e4 Qc5 14. Rac1 Rad8 15. Nd5 Bxd5 16. exd5 Rfe8 17. Ng3 Bh6 18. Rcd1 Nxd5 19. Qe4 Nf4 20. Qxb7 Qa5 21. Qf3 f5 22. Bc3 Qc5 23. Rxd8 Rxd8 24. Rd1 Rxd1+ 25. Qxd1 Qc6 26. f3 Qd6 27. Qxd6 cxd6 28. b4 d5 29. b5 dxc4 30. bxa6 Nd5 31. Kf2 Be3+ 32. Ke2 Bd4 33. Bxd4 exd4 34. Kd2 Kf7 35. a7 Nb6 36. Ne2 d3 37. Nf4 Ke8 38. Nd5 Na8 39. Ne3 Kd7 40. Nxc4 Kc6 41. a4 Kb7 42. Kxd3 Kxa7 43. g3 Kb7 44. Kd4 Kc6 45. f4 1-0

In this game, my highest rated opponent, Aydin Satici (2477) complimented me for saving the game with a knight sacrifice on move 35. I could only accept the compliment with guilt, as I have no way of knowing whether I would have found this subtle saving move, and I’m very surprised that Fritz found it (on an important side note, I wish to apologize to Aydin and all my other opponents for not really playing them. Some of these looked like very interesting games, and I would have loved to have participated!).

ICCF WC/65

Satici (2477) - Berch

1. c4 b6 2. Nc3 Bb7 3. e4 e6 4. g3 f5 5. Bg2 Nf6 6. d3 fxe4 7. Nge2 Bc5 8. O-O O-O 9. Bg5 Be7 10. dxe4 Ng4 11. Bxe7 Qxe7 12. b3 Nc6 13. Nf4 Nge5 14. Nh3 d6 15. f4 Nf7 16. Qd2 Rae8 17. Rad1 a6 18. Nf2 Qd8 19. Ng4 e5 20. f5 Bc8 21. Ne3 Nd4 22. Ne2 Nxe2+ 23. Qxe2 Ng5 24. b4 Bd7 25. b5 Bc8 26. h4 Nf7 27. f6 axb5 28. cxb5 Be6 29. Nf5 Bxf5 30. fxg7 Kxg7 31. exf5 Qf6 32. Qh5 e4 33. g4 e3 34. g5 Qe5 35. Qe2

Nxg5 36. hxg5 Rxf5 37. Rxf5 Qxf5 38. Rd4 Re5 39. Rc4 Re7 40. a4 d5 41. Rd4 Re5 42. Rg4 Re7 43. Bf3 Qe5 44. Kg2 Rf7 45. Qc2 1/2-1/2

Now, on to two less happy notes (if you’re rooting for the computer!). First, here is a draw that forced by one of the lower-rated players in a poisoned pawn Najdorf Sicilian.

ICCF WC XII/65

Stroeher (2029) - Berch

1. e4 c5 2. Nf3 d6 3. d4 cxd4 4. Nxd4 Nf6 5. Nc3 a6 6. Bg5 e6 7. f4 Qb6 8. Qd2 Qxb2 9. Rb1 Qa3 10. f5 Nc6 11. fxe6 fxe6 12. Nxc6 bxc6 13. e5 dxe5 14. Bxf6 gxf6 15. Ne4 Be7 16. Be2 h5 17. Rb3 Qa4 18. Nxf6 Bxf6 19. c4 Ra7 20. O-O Rd7 21. Qe3 Rf7 22. Rxf6 Rxf6 23. Qe5 Qxa2 24. Qxf6 Qxb3 25. Qxh8+ Kd7 26. Qd4+ Ke7 27. Bxh5 Qb4 28. Kf2 a5 29. Qg7+ Kd8 30. Qd4+ Ke7 1/2-1/2

This last game was one where one of the best players Fritz played made short work him with the Black pieces (as he probably would have done to me as well).

ICCF WC12/065

Berch - Hagstroem (2461)

1. d4 Nf6 2. c4 e6 3. Nf3 Bb4+ 4. Nbd2 b6 5. a3 Bxd2+ 6. Bxd2 Bb7 7. Bg5 d6 8. e3 Nbd7 9. Bd3 h6 10. Bh4 Qe7 11. Qc2 g5 12. Bg3 h5 13. h4 g4 14. Ng5 Bxg2 15. Rg1 Bb7 16. O-O-O O-O-O 17. b4 Kb8 18. Qc3 Rhg8 19. Bc2 Ka8 20. e4 Rxg5 21. hxg5 Nxe4 22. Bxe4 Bxe4 23. Qe3 Bf3 24. Rde1 Rg8 25. d5 e5 26. Bh4 Qd8 27. Qd3 e4 28. Qb3 Ne5 29. Kc2 Ng6 30. Bg3 h4 31. Bh2 Qxg5 0-1

There were a variety of other games, including a few where Fritz capitalized on a blunder, some endgames where Fritz was able to calculate well enough to hold a pawn down, a middle game position where moves that Fritz calculated as best by far eventually led to disaster, and a subtle endgame where Fritz "lost the thread." The games presented above, however, are (to me) the most interesting.