Difference between revisions of "Giraffe"

From Chessprogramming wiki
Jump to: navigation, search
(Created page with "'''Home * Engines * Giraffe''' FILE:The Burning Giraffe.jpg|border|right|thumb|240px|link=https://en.wikipedia.org/wiki/File:The_Burning_Giraffe.jpg| [[Ar...")
 
Line 7: Line 7:
  
 
=Description=  
 
=Description=  
Giraffe's [[Evaluation function|evaluation function]] is a [[Neural Networks#Deep|deep neural network]] trained by [[Temporal Difference Learning#TDLeaf|TDLeaf]] <ref>[[Jonathan Baxter]], [[Andrew Tridgell]], [[Lex Weaver]] ('''1998'''). ''TDLeaf(lambda): Combining Temporal Difference Learning with Game-Tree Search''. [https://www.chatbots.org/journal/australian_journal_of_intelligent_information_processing_systems/ Australian Journal of Intelligent Information Processing Systems], Vol. 5 No. 1, [http://arxiv.org/abs/cs/9901001 arXiv:cs/9901001]</ref> . Its feature representation includes a map of [[Static Exchange Evaluation|static exchange evaluations]] for all squares and sides <ref>[http://www.talkchess.com/forum/viewtopic.php?t=57045 SEE Map] by [[Matthew Lai]], [[CCC]], July 20, 2015</ref> , a structure already proposed by [[Russell M. Church]] and [[Kenneth W. Church]] in ''Plans, Goals, and Search Strategies for the Selection of a Move in Chess'' <ref>[[Russell M. Church]], [[Kenneth W. Church]] ('''1977'''). ''Plans, Goals, and Search Strategies for the Selection of a Move in Chess''. [[Chess Skill in Man and Machine]]</ref> . Probability-based evaluation [[Score|scores]] are not in [[Centipawns|centipawns]] nor linear to [[Material|material]] , and span a +-10,000 range, with [[Checkmate#MateScore|mate scores]] of +- 30,000. The [[Search|search]] recently changed from traditional depth-based [[Iterative Deepening|iterative deepening]] to assigning number of nodes (or time) to child nodes <ref>[http://www.talkchess.com/forum/viewtopic.php?t=57092 Time assignment to children] by [[Matthew Lai]], [[CCC]], July 26, 2015</ref> . Node budget allocation will also become [[Neural Networks|neural network]] based.
+
Giraffe's [[Evaluation Function|evaluation function]] is a [[Neural Networks#Deep|deep neural network]] trained by [[Temporal Difference Learning#TDLeaf|TDLeaf]] <ref>[[Jonathan Baxter]], [[Andrew Tridgell]], [[Lex Weaver]] ('''1998'''). ''TDLeaf(lambda): Combining Temporal Difference Learning with Game-Tree Search''. [https://www.chatbots.org/journal/australian_journal_of_intelligent_information_processing_systems/ Australian Journal of Intelligent Information Processing Systems], Vol. 5 No. 1, [http://arxiv.org/abs/cs/9901001 arXiv:cs/9901001]</ref> . Its feature representation includes a map of [[Static Exchange Evaluation|static exchange evaluations]] for all squares and sides <ref>[http://www.talkchess.com/forum/viewtopic.php?t=57045 SEE Map] by [[Matthew Lai]], [[CCC]], July 20, 2015</ref> , a structure already proposed by [[Russell M. Church]] and [[Kenneth W. Church]] in ''Plans, Goals, and Search Strategies for the Selection of a Move in Chess'' <ref>[[Russell M. Church]], [[Kenneth W. Church]] ('''1977'''). ''Plans, Goals, and Search Strategies for the Selection of a Move in Chess''. [[Chess Skill in Man and Machine]]</ref> . Probability-based evaluation [[Score|scores]] are not in [[Centipawns|centipawns]] nor linear to [[Material|material]] , and span a +-10,000 range, with [[Checkmate#MateScore|mate scores]] of +- 30,000. The [[Search|search]] recently changed from traditional depth-based [[Iterative Deepening|iterative deepening]] to assigning number of nodes (or time) to child nodes <ref>[http://www.talkchess.com/forum/viewtopic.php?t=57092 Time assignment to children] by [[Matthew Lai]], [[CCC]], July 26, 2015</ref> . Node budget allocation will also become [[Neural Networks|neural network]] based.
  
 
=See also=  
 
=See also=  

Revision as of 22:44, 3 June 2018

Home * Engines * Giraffe

Giraffe,
an experimental open source chess engine by Matthew Lai under the GNU General Public License, compliant to the Chess Engine Communication Protocol, written in C++11 and based on deep learning, which is topic of his Master's thesis in August 2015 [2] [3] . Giraffe uses the Eigen linear algebra library [4] , and Pradyumna Kannan's magic move generator [5] [6]. As employee of Google DeepMind, Matthew Lai announced the discontinuation of the Giraffe project in January 2016 [7].

Description

Giraffe's evaluation function is a deep neural network trained by TDLeaf [8] . Its feature representation includes a map of static exchange evaluations for all squares and sides [9] , a structure already proposed by Russell M. Church and Kenneth W. Church in Plans, Goals, and Search Strategies for the Selection of a Move in Chess [10] . Probability-based evaluation scores are not in centipawns nor linear to material , and span a +-10,000 range, with mate scores of +- 30,000. The search recently changed from traditional depth-based iterative deepening to assigning number of nodes (or time) to child nodes [11] . Node budget allocation will also become neural network based.

See also

Publications

Forum Posts

2015

2016

2017...

Re: Is AlphaGo approach unsuitable to chess? by Peter Österlund, CCC, May 31, 2017 » Texel

External Links

Chess Engine

Misc

Giraffe - Internal systems

References

Up one Level