Fortress,
an endgame drawing technique where a side usually behind in material sets up a zone of protection that the opponent cannot penetrate [1]. While fortresses are quite common in (late) endgames, f.i. in Wrong color Bishop and rook pawn, fortresses may rarely occur in the (late) middlegame, characterized by long fixed pawn chains with up to one or two open files but all points of penetration well defended by the weaker side. Chess programs without the implemetation of special knowledge typically fail to recognize fortresses and seem to claim a winning advantage, although they are not able to achieve the win against adequate defence [2].
Fortress detection might be triggered in evaluation or interior node recognizers if certain material configurations along with the winning advantage, and pawn structure properties occur. Further pattern matching considering relevant features of fortress positions or fuzzier pattern recognition is then used to scale the winning advantage towards the draw score zero - the fuzzier and possibly more error-prone the heuristics, the less the draw scaling. However, due to search versus knowledge trade-off, and all that generalization and tuning problems involved with adding knowledge and noise, practical playing strength versus usefulness in analysis is an issue.
Sample
A pattern of rook vs queen fortress in the endgame was given by Lyudmil Tsvetkov in his 2017 The Secret of Chess[4]:
One side having just one queen, and at most a single pawn, with the other having just one rook, and at least one pawn, if the queen side has no passers, the rook is protected by a pawn on the g or b files on its 2nd rank, the king of the rook side is adjacent to that pawn, the pawn of the queen side is on files h,f,a or c, and the king of the queen side is not past its 5th rank.
Search
Another technique of fortress detection as proposed by Matej Guid and Ivan Bratko in 2012 [5] is related to search, in particular iterative deepening along with "no progress" indication - that is the root evaluation or low search depths already indicate a winning score, which is not increasing in consecutive iterations. A few programs start scaling scores toward zero if the halfmove clock already exceeds values far below enforcing the 50 move rule, often triggered by conditions mentioned above. Matej Guid and Ivan Bratko further mention the possibility of Monte-Carlo Tree Search to detect fortresses.
an endgame drawing technique where a side usually behind in material sets up a zone of protection that the opponent cannot penetrate [1]. While fortresses are quite common in (late) endgames, f.i. in Wrong color Bishop and rook pawn, fortresses may rarely occur in the (late) middlegame, characterized by long fixed pawn chains with up to one or two open files but all points of penetration well defended by the weaker side. Chess programs without the implemetation of special knowledge typically fail to recognize fortresses and seem to claim a winning advantage, although they are not able to achieve the win against adequate defence [2].
Table of Contents
Fortress Detection
Evaluation
Fortress detection might be triggered in evaluation or interior node recognizers if certain material configurations along with the winning advantage, and pawn structure properties occur. Further pattern matching considering relevant features of fortress positions or fuzzier pattern recognition is then used to scale the winning advantage towards the draw score zero - the fuzzier and possibly more error-prone the heuristics, the less the draw scaling. However, due to search versus knowledge trade-off, and all that generalization and tuning problems involved with adding knowledge and noise, practical playing strength versus usefulness in analysis is an issue.Sample
A pattern of rook vs queen fortress in the endgame was given by Lyudmil Tsvetkov in his 2017 The Secret of Chess [4]:Search
Another technique of fortress detection as proposed by Matej Guid and Ivan Bratko in 2012 [5] is related to search, in particular iterative deepening along with "no progress" indication - that is the root evaluation or low search depths already indicate a winning score, which is not increasing in consecutive iterations. A few programs start scaling scores toward zero if the halfmove clock already exceeds values far below enforcing the 50 move rule, often triggered by conditions mentioned above. Matej Guid and Ivan Bratko further mention the possibility of Monte-Carlo Tree Search to detect fortresses.See also
Selected Publications
Forum Posts
2000 ...
2005 ...
2010 ...
2015 ...
External Links
On Topic
Misc
References
What links here?
Up one Level