What are the most important 6 man TBs to have

[jdart]

It can be dangerous to have the pawn TBs but not the ones that result from promotion of pawns - after one or more promotions you can get dumped into a different ending that you are all of a sudden having to use search to solve. Some are trivial but others are not.

[kasinp]

Interesting. How were these figures calculated?

Chessbase allows you to search for specific material on the board.
A while ago I did an exhaustive analysis of a large comp-comp database. IIRC I extracted all positions where a 5-man endgame occurred for at least 2 plies.

Percentages refer to the total number of 5-man endgames in the database. The krppkr endgame was responsible for (roughly) every seventh such case.

PK

[kurtneumann]

looks acceptable. especially KP endings

I would say, minimally, get these:

KPPKPP (great for king and pawn endings, of course)
KQPKPP (for when one side just barely outpromotes the other)

KQPKQP (equal material with 1 pawn each)
KRPKRP
KBPKBP
KNPKNP

KQPKRP (major vs. major with 1 pawn each)
KBPKNP (minor vs. minor with 1 pawn each)

Whether you want this one (sometimes the knights win) is a coin toss:

KNNKPP

Just my opinion, of course.

[LiquidNitrogenOverclocker]

It can be dangerous to have the pawn TBs but not the ones that result from promotion of pawns - after one or more promotions you can get dumped into a different ending that you are all of a sudden having to use search to solve. Some are trivial but others are not.

I think you are talking about over-the-board play. I am talking about probing the TB while the program is searching.

As soon as you hit a leaf node, such as KPPKPP, the tablebase returns the score immediately. There is no move generation, so any potential pawn promotion (which might result in a tablebase you don't have) does not occur. The tree is reduced, fewer nodes, with better information.

The idea is to guide the program to a winning position from a great distance away. The program can mostly likely still win, although much less efficiently, if it is missing a supporting TB, and the position is reached OTB.

[BubbaTough]

It can be dangerous to have the pawn TBs but not the ones that result from promotion of pawns - after one or more promotions you can get dumped into a different ending that you are all of a sudden having to use search to solve. Some are trivial but others are not.

I think you are talking about over-the-board play. I am talking about probing the TB while the program is searching.

As soon as you hit a leaf node, such as KPPKPP, the tablebase returns the score immediately. There is no move generation, so any potential pawn promotion (which might result in a tablebase you don't have) does not occur. The tree is reduced, fewer nodes, with better information.

The idea is to guide the program to a winning position from a great distance away. The program can mostly likely still win, although much less efficiently, if it is missing a supporting TB, and the position is reached OTB.

Jon is right. I have seen it in action in many games. For example:
You have KPPKPP but not KPQKPQ.
During the search, the computer could decide to go into a won KPPKPP endgame instead of taking a trivial win (say a rook up endgame) because the tablebase promises a forced mate, while the won endgame is only +8.32. So the program goes into the endgame, and plays out the winning moves, which involves both sides promoting, turning it into a won KPQKPQ endgame. But wait! You don't have that tablebase, and without it its very very hard to win! So your program ends up drawing its totally won game.

I know this scenario sounds far fetched, but I have seen it happen a number of times.

-Sam

[LiquidNitrogenOverclocker]

Jon is right. I have seen it in action in many games. For example:
You have KPPKPP but not KPQKPQ.
During the search, the computer could decide to go into a won KPPKPP endgame instead of taking a trivial win (say a rook up endgame) because the tablebase promises a forced mate, while the won endgame is only +8.32. So the program goes into the endgame, and plays out the winning moves, which involves both sides promoting, turning it into a won KPQKPQ endgame. But wait! You don't have that tablebase, and without it its very very hard to win! So your program ends up drawing its totally won game.

I know this scenario sounds far fetched, but I have seen it happen a number of times.

-Sam

Except I already specified an entire set of TBs in my original reply, and it did include the KQPKQP set.

Of course one TB is not a Panacea.

[BubbaTough]

Jon is right. I have seen it in action in many games. For example:
You have KPPKPP but not KPQKPQ.
During the search, the computer could decide to go into a won KPPKPP endgame instead of taking a trivial win (say a rook up endgame) because the tablebase promises a forced mate, while the won endgame is only +8.32. So the program goes into the endgame, and plays out the winning moves, which involves both sides promoting, turning it into a won KPQKPQ endgame. But wait! You don't have that tablebase, and without it its very very hard to win! So your program ends up drawing its totally won game.

I know this scenario sounds far fetched, but I have seen it happen a number of times.

-Sam

Except I already specified an entire set of TBs in my original reply, and it did include the KQPKQP set.

Of course one TB is not a Panacea.

Well then, in addition to supporting Jon's generic point, my comments support the quality of your list . I don't think people have to worry about under-promotion too much.

-Sam