Fail-high reductions (FHR),
as proposed and examined by Rainer Feldmann , and implemented in the program Zugzwang, are based on the Null Move Observation (NMO) - similar to the Null Move Heuristic (NMH). The idea is to search to shallower depth at positions that are seemingly quiet and where the side to move has established a substantial advantage, according to a static evaluation. Apart from the evaluation function, the heuristic requires an additional function that returns a value indicating threats against the side to move and therefor the quietness of a position.
FHR is applied at expected Fail-high or Cut-Nodes recursively inside a NegaScout-framework. Feldmann's implementation did no re-search with the original depth, if the shallow search didn't confirm the fail high. While NMH relies on a dynamic search, but cuts the whole subtree - FHR uses a static threat detection. FHR is not as vulnerable as NMH in situations, where the NMO fails - namely in zugzwang.
eval, threat := evaluate(...); if ( eval - threat >= beta && alpha == beta - 1) reduce depth by one
- Double Null Move
- Late Move Reductions
- Null Move Pruning
- Null Move Reductions
- Zugzwang (Program)
- Rainer Feldmann (1997). Fail-High Reductions. Advances in Computer Chess 8, available as pdf from CiteSeerX
- Rainer Feldmann, Burkhard Monien (1998). Selective Game Tree Search on a Cray T3E. ps
- Yngvi Björnsson, Tony Marsland (2000). Selective Depth-First Search Methods. Games in AI Research (eds. Jaap van den Herik and Hiroyuki Iida), pp. 31-45. Universiteit Maastricht, Maastricht, The Netherlands. ISBN 90-621-6416-1. pdf preprint
- Fail High reductions by Jon Dart, rgcc, October 7, 1996
- Fail high reductions by Russell Reagan, CCC, July 01, 2003
- Fail High Reductions - question about researches by milix, CCC, November 11, 2004