Author: Robert Hyatt
Date: 05:33:02 05/14/98
Go up one level in this thread
On May 14, 1998 at 01:07:29, Don Dailey wrote: >On May 13, 1998 at 21:59:41, Howard Exner wrote: > >>Some of the games I've managed to win against computers, and these are >>few in comparison to the many defeats I've suffered, contain the theme >>of the "trapped piece". Common themes are Knights trapped in the corner >>and basically pieces that have their mobility stripped. >> >>To illustrate here is a line that I've played from the Crafty vs >>Rebel game from KKup2. >> >>[FEN "8/2N1R3/7p/R4r2/3pkr1P/1P6/2P5/2K2b2 b - - 0 1"] >> >>1... Kf3 2. Rxf5 Rxf5 3. Rd7 Ke4 4. Kd2 Rh5 5. Ne6 Ke5 6. Nxd4 Rxh4 7. >>Nf3+ Ke6 8. Nxh4 Kxd7 9. Nf5 Bh3 10. Nxh6 Be6 11. c4 Kc6 12. Kc3 Kc5 13. >>b4+ Kc6 14. Kd4 Kb6 15. Kd3 Kc6 16. Kc3 Kb6 17. Kd4 Kc6 * >> >>Granted it was played rather quickly so I doubt very much that this >>will transpire in the real game. I took black's side here and had >>Rebel 8 play for white. After the move 10. Nxh6 Rebel thinks its >>position is quite good, but the move Be6 traps the Knight. This kind >>of position is nothing new to the readers here but it made me wonder >>how programmers deal with this type of thing. How can positions were >>pieces get trapped be avoided? I see no way for white to make any >>progress from the above line despite the material plus. Black simply >>has its king prevent the advance of the white pawns. >> >>Would any program play the move 9 ... Bh3 so as to trap the Knight >>if Nxh5? > >8/3k4/4b2N/8/8/1P6/2PK4/8 w - - > >Whites advantage is pretty small in this position, I see a couple tries >but they lead to draws. But the point is having some concept of a >trapped piece. It would not be hard putting this particular pattern >in a chess program but like many other chess principles it becomes >difficult to come up with proper scoring. Pehaps in similar positions >the knight can be extracted. In this position extracting the knight >costs a pawn (and half a point.) > >The problem isn't so much detecting a pin piece (which of course is >not easy either but would be useful to do) but figuring out how it >relates to all the other moves and the rest of the game. Maybe >it's ok, maybe it's not. As soon as a human see's this position, >he immediately begins to reason about it in a way that computers >do not currently do (as far as I know.) This becomes critical >and powerful information for the human. Suddenly things occur >to him, plans to extract the knight, he knows exactly where the >king must be to extract it, or he reasons that the bishop might >be forced to defend thereby freeing the knight. And he considers >the repercussions of all of this and knows how to integrate the >whole process into a serious of quality moves. But the computer >does not even really know the knight is trapped! As far as the >computer is concerned, h6 is the best square for the knight since >the other squares lose it. And that's all it cares about. It >of course knows other squares are better but it is completely >isolated from this knowledge because the search says it's not >relevant here! Cool and efficient in a way. Do not worry about >something you cannot have! > >Several years ago, Rexchess (on route to winning a human tournament) >captured a rook on a8 getting its knight trapped. It was the classical >computer "blunder" because material had been sacrafice in order >to "win" the rook. To save the knight, Rex suddenly woke up and >began a "reign of terror" on it's poor unsuspecting opponent. It >began a beautiful and furious attack which quickly won >the game. I recognized immediately that it was the horizon affect >kicking in with a vengance and vowed right then and there I would >fix this problem right away. > >But after looking at the game later, >it became quite clear that the whole plan was quite sound! The >rook was completely takeable although most of us watching >laughed (except for me, my heart broke) when the computer played >this "silly move." This game made me see the horizon affect problem >from another perspective. In a sense there is no such thing as >the horizon affect, or if you prefer it's an artificial concept >we made up to explain the way we humans look at it. If this is >unclear, let me state it another way. Every single move searched >by a chess program (except ones ending game theoreticially) are >horizon affect moves, but we only use that terminology when WE >see something the computer doesn't. We never call it horizon >affect when the computer gets gradually outplayed. > >I was fascinated by the fact that the computers extreme naivety >is what caused it to win. It's like the fighter who doesn't >know he's supposed to lose and so goes on to win. > >I've discovered over time that the computer seems to make this >decision correctly most of the time. This is due to greater >search depths. I remember Bob Hyatt telling me that early Blitz >program had this rule and it was useful, later he felt like it >hurt the program. I can believe this. > >So many programs do have trapped piece code in them, but I really >doubt they are really integrated very well into the overall >strategic decision making, they are probably more prophylactic >in nature. > >- Don I only fix things that are obviously "broken". The most common trapped piece has *always* been Bxa2, Bxh2, Bxh7 and Bxa7 winning a pawn, but getting the bishop trapped. I don't *ever* make these mistakes as I have special code. I handle the knight case (giving up a piece so that I can play Nxa8 for example) by simply having a huge negative penalty for a knight in the corner. If it can rip the rook *and* get the knight out, all is well, otherwise, it gets the penalty and refuses to enter into that line. Other cases of trapped/pinned pieces I don't handle specifically, depending on deep searches to work those out...
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.