25,161
edits
Changes
no edit summary
a [[Best-First|best-first search]] algorithm first described by [[David McAllester]] based on [[Conspiracy Numbers]] of the [[Root|root]] <ref>[[David McAllester]] ('''1985'''). ''A New Procedure for Growing Minimax Trees''. Technical Report, Artificial Intelligence Laboratory, [[Massachusetts Institute of Technology|MIT]]</ref>.
[[Search Tree|Trees]] are grown in [[Memory|memory]] - in an often deep and narrow way - that maximizes the conspiracy required to change the root value.
The phases of the best-first search procedure are '''Selection''' of a [[Leaf Node|leaf node]], '''Expansion''' and '''Evaluation''' of that leaf, and to '''Back-up''' the result of that evaluation back to the root. =CNS=CNS maintains a range of possible values and keeps expanding the tree until a certain degree of confidence is reached. The confidence is measured by the width of a possible values’ range W and a minimum value for conspiracy numbers T. The purpose of the search is to raise the conspiracy numbers of unlikely values to greater than T in order to reduce the range of possible values to below W. At each turn, CNS tries to disprove either the highest or lowest possible value, which has the highest conspiracy numbers, by expanding one of its conspirators. Then, it recalculates conspiracy numbers and repeats the process until the desired confidence is obtained <ref>[[Quang Vu]], [[Taichi Ishitobi]], [[Jean-Christophe Terrillon]], [[Hiroyuki Iida]] ('''2016'''). ''Using Conspiracy Numbers for Improving Move Selection in Minimax Game-Tree Search''. [http://www.icaart.org/?y=2016 ICAART 2016], [https://pdfs.semanticscholar.org/1bcf/bd2047bc1d74affda11bf2007cac442dd6f4.pdf pdf]</ref>.Since the number of conspiracy numbers per [[Node|node]] depends on the number of possible evaluation values, fine grained evaluation, necessary for good positional play, yields in inefficient and unstable CNSsearch. Further, due to the lack of [[Bound|bounds]], a final [[Alpha-Beta|alpha-beta]] [[Quiescence Search|quiescence search]] in early CNS implementations was quite expensivedue to the lack of [[Bound|bounds]].
<span id="CCNS"></span>
=Controlled Conspiracy Number SearchCCNS=
[[Ulf Lorenz|Ulf Lorenz']] and [[Valentin Rottmann|Valentin Rottmann's]] et al. proposed improvements dubbed '''Controlled Conspiracy Number Search''' (CCNS) <ref>[[Ulf Lorenz]], [[Valentin Rottmann]], [[Rainer Feldmann]], [[Peter Mysliwietz]] ('''1995'''). ''Controlled Conspiracy-Number Search.'' [[ICGA Journal#18_3|ICCA Journal, Vol. 18, No. 3]]</ref> address the drawbacks of CNS by introducing general '''CN Targets''' and '''Extended Conspiracy Numbers'''. CN targets (security demands) are splitted over the successors in order to inform each node about the goal of its examination.
Extended Conspiracy Numbers of the root are defined as the least number of leaf nodes that must change their value in order to change the decision at the root to another move.
<span id="PCCNS"></span>
=PCCNS= '''Parallel Controlled Conspiracy Number Search= The parallelization procedure ''' (PCCNS) aims at a dynamic distribution of the game tree, initiating a worker/employer relationship along with a sophisticated splitting heuristic of CN targets.The [[Stack|stack]], used to keep the nodes of the best-first search, could be manipulated from outside in order to share work and integrate results from other processors <ref>[[Ulf Lorenz]], [[Valentin Rottmann]] ('''1996'''). ''Parallel Controlled Conspiracy-Number Search.'' [[Advances in Computer Chess 8]]</ref>.
=See also=
