Hello everyone, pleased to meet you all. This is my first post here.
I've played casual chess my whole life (no rating, but maybe 2000-ish if I had to guess), and I've recently discovered tablebases. They're amazing! I've been really working on my endgame skills, and have long been fascinated with Queen vs Rook. I've done tons of work on the ending, with spreadsheets. My question probably differs from what most people are after in these forums. I'm not looking to create a tablebase for a computer to play; I'm looking for a tablebase of just Queen vs Rook that I can MEMORIZE.
I know, it's impossible to memorize that much. But I still want to work on it for fun. Now, a lot of memorizing would be eliminated by mentally combining similar positions and classifying them as the same (such as back row mates, etc), and also by only focusing on the fastest route to EITHER mate OR capturing the rook (I don't need the tablebase to continue after the rook dies. I think it's called DTZ or DTR or DTR50 or something like that, I still don't understand the difference between them all - and yes, I care about the 50 move rule because in tournament play, it will apply). I would also like to eliminate redundant lines. For example, if there's a mate in 12 and a mate in 14, but the one in 14 moves involves moves that make no actual progress, then I wouldn't even want it to show up.
Also, I only care about learning how to win it with white (For now. Later I may work on defending as black). Also, if there is a continuation that wins in 10 moves, but has a dozen variations, and another one wins in 11 yet it's all forced, I'd rather have the one that's forced, since it's less memorizing. I don't know if I'm being clear or not, it's hard to describe what I'd like to create (actually I don't want to create it, I just want to own it, as a tool, to truly learn the ending). I guess I'd like to see a complete tablebase that wins for white using the most efficient moves to mate or capture, yet in its easiest form for a human to learn.
Maybe it would make more sense to have something that shows all the wins in 1 move, then all the ones that can reach THOSE positions in 1 move, then all the positions that can reach THOSE in 1 move... And then I can study them starting with simplest and working my way back. I think that's how tablebases are created anyway, isn't it? Anyway...
Is it hopeless for a human to get to the point that he/she can look at any position on the board involving Queen vs Rook and be able to deliver checkmate or win the rook with say, 1 minute on the clock, doing each move without hesitation? Or is it an impossible dream? How would you guys learn it?
Thanks!
KQ vs KR, human
-
guyhaw
- Posts: 489
- Joined: Sat Jan 21, 2006 10:43 am
- Sign-up code: 10159
- Location: Reading, UK
- Contact:
Re: KQ vs KR, human
The DTM (Depth to Mate) EGT is publicly available. The DTC (Depth to Conversion) is probably better for learning how to win but is not so available. Neither recognise the 50-move rule but you should not need to for KQKR. http://chess.jaet.org/endings/ gives both DTC and DTM depths for KQKR positions.
John Nunn's 'Secrets of Pawnless Endings', now in Edition 2, devotes Chapter 3 (pp49-69, positions 63-94) to KQKR: highly recommended. He agrees that the EGTs underline how difficult it is to win, even though it has been known that this is a 'general win' for over 200 years. The deepest wins have DTC=31 and the two Browne-BELLE demonstrator games showed that even a GM struggles to win a 'DTC=31' position in 50 moves.
Browne managed 50% (just) after studying BELLE's output from the first game for some weeks: BELLE played DTC-optimal moves but sometimes did not pick the DTC-optimal move giving Browne the toughest challenge.
JN highlights these basic ideas:
1) Distance wK and wQ from threats by the bR
2) Avoid potential stalemates
3) White wins by forking K & R, or by first driving the bK to the edge of the board
4) Black defends by trying to cut off the wK from the attack with its Rook, see '2nd/3rd/4th-rank defence'
5) So, part of 'driving to the edge' involves lifting the blockade by the bR, relatively easy for the 4th-rank defence
6) The 'Philidor position' is notable: Kc6 Qa5 kb8 rb7 ... and appeared in a very, very long game won by Kosteniuk recently (see Krabbe Chess Diary)
The Kosteniuk game was (?!) actually at a chess/music event celebrating Philidor, and - as they had agreed not to involve the Arbiter, the KQKR phase went on for much more than 50 moves.
I don't have any stats for frequency of occurence of this endgame, positions won/drawn when they should (not) have been won/drawn, etc. JN includes didactic positions, studies and extracts from 11 games JN rates a win of 12m as 'relatively simple', 15m deep is 'medium difficulty' and 20m deep is 'difficult'.
g
John Nunn's 'Secrets of Pawnless Endings', now in Edition 2, devotes Chapter 3 (pp49-69, positions 63-94) to KQKR: highly recommended. He agrees that the EGTs underline how difficult it is to win, even though it has been known that this is a 'general win' for over 200 years. The deepest wins have DTC=31 and the two Browne-BELLE demonstrator games showed that even a GM struggles to win a 'DTC=31' position in 50 moves.
Browne managed 50% (just) after studying BELLE's output from the first game for some weeks: BELLE played DTC-optimal moves but sometimes did not pick the DTC-optimal move giving Browne the toughest challenge.
JN highlights these basic ideas:
1) Distance wK and wQ from threats by the bR
2) Avoid potential stalemates
3) White wins by forking K & R, or by first driving the bK to the edge of the board
4) Black defends by trying to cut off the wK from the attack with its Rook, see '2nd/3rd/4th-rank defence'
5) So, part of 'driving to the edge' involves lifting the blockade by the bR, relatively easy for the 4th-rank defence
6) The 'Philidor position' is notable: Kc6 Qa5 kb8 rb7 ... and appeared in a very, very long game won by Kosteniuk recently (see Krabbe Chess Diary)
The Kosteniuk game was (?!) actually at a chess/music event celebrating Philidor, and - as they had agreed not to involve the Arbiter, the KQKR phase went on for much more than 50 moves.
I don't have any stats for frequency of occurence of this endgame, positions won/drawn when they should (not) have been won/drawn, etc. JN includes didactic positions, studies and extracts from 11 games JN rates a win of 12m as 'relatively simple', 15m deep is 'medium difficulty' and 20m deep is 'difficult'.
g
Re: KQ vs KR, human
I would look for a book in which the author explains the basic techniques and classifies the most important patterns.simaton wrote:How would you guys learn it?
I'm not aware of any such book or explanation for KQKR (but I never looked for one either).
-
simaton
Re: KQ vs KR, human
Thank you so much for that link! Yes, I'm aware of the books and the general principles, but a side-by-side DTC and DTM is exactly what I was looking for. This will greatly speed up what I've been doing (using a DTM and manually turning it into a DTC using a spreadsheet).
-
h.g.muller
- Posts: 223
- Joined: Mon Feb 19, 2007 8:24 am
- Sign-up code: 0
- Location: Amsterdam
- Contact:
Re: KQ vs KR, human
You touch on an interesting subject, which I often wondered about. The DTM or DTC are an absolute measure for the winning distance, but very erratically related to the actual position. This makes it impossible for a Human to remember them, and makes it very difficult for a computer program to develop 'recognizers": small evaluation routines for a given end-game that would replace the table lookup.
Even if th purpose is perfect play, the absolute DTx is 'overkill', and a relative progress measure could aleady be sufficient. You could for instance build a tablebase in the usual way, except that you don't assign the predecessors of a won or lost position with DTx=N not DTx = N+1, but DTx = N+r, where r is some randomly chosen positive number r. If you make sure in the verification process that r is chosen large eough to exceed the DTx of all successors that are already decided, you get an EGT that would cause identical play to a conventional one (i.e. always pick the shortest line for the attacker or longest line for the defender). The only thing that would be missing is that you cannot predict from the progress measure how long exacty it is going to take t effect the win. The relation between true DTx and this new distance measure does not even have to be globally monotonic, as long as it is mnotonic along all possible shortest winning paths. The relative ordering of positions that are on non-intersecting paths, is immaterial.
This freedom could be used to group together positions with the similar (Human-recogniable) chractristics, that would have different absolute DTx in an exact EGT. Such positions would than get the same relative progress measure. This corresponds much more to how Humans remember end-game techniques: Define an ordedered list of sub-goals. First achieve A, from there on force B from there on C. If from some positions of the A group, you can force condition C directly, you simply skip B. E.g. in KBNK, first achieve opposition in the center, then drive the bare King to the edge, then trap him in the correct corner with K + one piece, then bring the second piece in a position to checkmate, and finally checkmate. The EGTs with absolute DTx assign different DTx to positions in the same group, making them hard to remember. Not having to worry about details of the position, which do affect the mating distance because you might have to interject an extra move later to reach a sub-goal that would be automatically satisfied in another line (e.g. distance your Bishop from the bare King, because it as to close to move to the diagnal you want it on to go there unattacked).
If you drop the requirement that play should be perfect, and replace it by the weaker condition that the progress measure contains no loops (so that the conversion or mating goal will always be attained within 50 moves, despite the fact that the absolute DTx might go up in some parts of the path), you even get more freedom in assigning progress values to positions in a way that makes it easy to recognize them from characteristics of the position.
Even if th purpose is perfect play, the absolute DTx is 'overkill', and a relative progress measure could aleady be sufficient. You could for instance build a tablebase in the usual way, except that you don't assign the predecessors of a won or lost position with DTx=N not DTx = N+1, but DTx = N+r, where r is some randomly chosen positive number r. If you make sure in the verification process that r is chosen large eough to exceed the DTx of all successors that are already decided, you get an EGT that would cause identical play to a conventional one (i.e. always pick the shortest line for the attacker or longest line for the defender). The only thing that would be missing is that you cannot predict from the progress measure how long exacty it is going to take t effect the win. The relation between true DTx and this new distance measure does not even have to be globally monotonic, as long as it is mnotonic along all possible shortest winning paths. The relative ordering of positions that are on non-intersecting paths, is immaterial.
This freedom could be used to group together positions with the similar (Human-recogniable) chractristics, that would have different absolute DTx in an exact EGT. Such positions would than get the same relative progress measure. This corresponds much more to how Humans remember end-game techniques: Define an ordedered list of sub-goals. First achieve A, from there on force B from there on C. If from some positions of the A group, you can force condition C directly, you simply skip B. E.g. in KBNK, first achieve opposition in the center, then drive the bare King to the edge, then trap him in the correct corner with K + one piece, then bring the second piece in a position to checkmate, and finally checkmate. The EGTs with absolute DTx assign different DTx to positions in the same group, making them hard to remember. Not having to worry about details of the position, which do affect the mating distance because you might have to interject an extra move later to reach a sub-goal that would be automatically satisfied in another line (e.g. distance your Bishop from the bare King, because it as to close to move to the diagnal you want it on to go there unattacked).
If you drop the requirement that play should be perfect, and replace it by the weaker condition that the progress measure contains no loops (so that the conversion or mating goal will always be attained within 50 moves, despite the fact that the absolute DTx might go up in some parts of the path), you even get more freedom in assigning progress values to positions in a way that makes it easy to recognize them from characteristics of the position.
-
guyhaw
- Posts: 489
- Joined: Sat Jan 21, 2006 10:43 am
- Sign-up code: 10159
- Location: Reading, UK
- Contact:
Lattice of subgoals
Let us, for the moment, ignore the k-move rule which will really complicate things 
.
It is certainly true that if one could have, for each position, a defined loopless, sequence (i.e. lattice) of subgoals ending in mate which can be persistently (i.e. without subsequent regression) achieved, then 'depth' could be measured in terms to next subgoal. hgm provides a valuable reminder of that idea. 'Conversion' and 'zeroing move-count' are such goals, hence the DTC and DTZ metrics. 'Getting depth down to d' in whatever metric is also such a goal.
There have been some notable initiatives in this area, but none recently as far as I know. John Roycroft worked with (the now late) Donald Michie to define an approach to wins in KBBKN: Ken Thompson's EGT had just changed perceptions of KBBKN. AJR defined five phases of winning KBBKN but did not prove that if you are in phase 'n' then you necessarily have to go to phase 'n+1' next - so there is still work to do on KBBKN...
... If one wants to know what decisive positions P' depend on decisive positions P, then my Scorched Earth Algorithm (SEA) provides the answer. Consider the chess variant Chess(P) in which positions in P are deemed to be draws: calculate the EGT for Chess(P) and see which positions P' >= P have become draws. This approach can also be used to validate some technical aspects of Chess Studies, the original motivation for SEA, and obviously to guard against unwelcome repetition of positions.
JR Quinlan ['Google Scholar' Quinlan] used IDE (?) and the chess domain to demonstrate the induction of decision-trees to solve problems like 'winning'. But even KRK and KBNK are hard to crack. What is far easier is to get players with vision and an ability to communicate to tell us what the subgoals should be. Humans have an abstracting, inductive ability 'built in' to a degree, and good players have a lot of it in the chess domain. John Nunn often gives a winning line governed by more understandable rules rather than the DTM-minimaxing line. Anand won the other day with a clear (Bh3) rather than quick plan.
Here in KQKR, there would appear to be goals like 'fork, skewer or Chinese-skewer kr if possible', and 'constrain the bK to the back 3 (2, 1) rank(s)'.
http://portal.acm.org/citation.cfm?id=1137604 includes mention of KQKR.
g
. It is certainly true that if one could have, for each position, a defined loopless, sequence (i.e. lattice) of subgoals ending in mate which can be persistently (i.e. without subsequent regression) achieved, then 'depth' could be measured in terms to next subgoal. hgm provides a valuable reminder of that idea. 'Conversion' and 'zeroing move-count' are such goals, hence the DTC and DTZ metrics. 'Getting depth down to d' in whatever metric is also such a goal.
There have been some notable initiatives in this area, but none recently as far as I know. John Roycroft worked with (the now late) Donald Michie to define an approach to wins in KBBKN: Ken Thompson's EGT had just changed perceptions of KBBKN. AJR defined five phases of winning KBBKN but did not prove that if you are in phase 'n' then you necessarily have to go to phase 'n+1' next - so there is still work to do on KBBKN...
... If one wants to know what decisive positions P' depend on decisive positions P, then my Scorched Earth Algorithm (SEA) provides the answer. Consider the chess variant Chess(P) in which positions in P are deemed to be draws: calculate the EGT for Chess(P) and see which positions P' >= P have become draws. This approach can also be used to validate some technical aspects of Chess Studies, the original motivation for SEA, and obviously to guard against unwelcome repetition of positions.
JR Quinlan ['Google Scholar' Quinlan] used IDE (?) and the chess domain to demonstrate the induction of decision-trees to solve problems like 'winning'. But even KRK and KBNK are hard to crack. What is far easier is to get players with vision and an ability to communicate to tell us what the subgoals should be. Humans have an abstracting, inductive ability 'built in' to a degree, and good players have a lot of it in the chess domain. John Nunn often gives a winning line governed by more understandable rules rather than the DTM-minimaxing line. Anand won the other day with a clear (Bh3) rather than quick plan.
Here in KQKR, there would appear to be goals like 'fork, skewer or Chinese-skewer kr if possible', and 'constrain the bK to the back 3 (2, 1) rank(s)'.
http://portal.acm.org/citation.cfm?id=1137604 includes mention of KQKR.
g