Difference between revisions of "Floyd"

From Chessprogramming wiki
Jump to: navigation, search
(Created page with "'''Home * Engines * Floyd''' FILE:Hurricane Floyd 14 sept 1999 2030Z.jpg|border|right|thumb|Hurricane Floyd <ref>Hurricane Floyd near peak intensity on Se...")
 
Line 21: Line 21:
  
 
==Evaluation==  
 
==Evaluation==  
Floyd's [[Evaluation|evaluation]] employs a vector of feature and weight pairs to calculate a [[Score|score]] as [https://en.wikipedia.org/wiki/Weighted_sum_model weighted sum]. In conjunction with a draw model <ref>[https://github.com/kervinck/floyd/blob/master/Docs/drawModel.txt floyd/drawModel.txt at master · kervinck/floyd · GitHub]</ref> using [https://en.wikipedia.org/wiki/Sigmoid_function sigmoid functions], the score is mapped to [[Pawn Advantage, Win Percentage, and ELO|winning probabilities]], suited for [[Automated Tuning#LogisticRegression|logistic regression]] tuning.
+
Floyd's [[Evaluation|evaluation]] employs a vector of feature and weight pairs to calculate a [[Score|score]] as [https://en.wikipedia.org/wiki/Weighted_sum_model weighted sum]. In conjunction with a draw model <ref>[https://github.com/kervinck/floyd/blob/master/Docs/drawModel.txt floyd/drawModel.txt at master · kervinck/floyd · GitHub]</ref> using [https://en.wikipedia.org/wiki/Sigmoid_function sigmoid functions], the score is mapped to [[Pawn Advantage, Win Percentage, and Elo|winning probabilities]], suited for [[Automated Tuning#LogisticRegression|logistic regression]] tuning.
 
<pre>
 
<pre>
 
def evaluate(board, wiloVector, drawVector):
 
def evaluate(board, wiloVector, drawVector):

Revision as of 23:14, 18 June 2018

Home * Engines * Floyd

Hurricane Floyd [1]

Floyd,
an UCI compliant open source chess engine for study purposes and prototyping of new ideas by Marcel van Kervinck, written in C, and first released in October 2015 [2] with a permissive license [3] . Floyd can be build to run under Windows, Linux and Mac OS. Floyd had its over the board tournament debut at the IGT 2016 with a 50% score.

Description

Board Representation

Floyd uses an 8x8 Board, agnostic to square indexing, in the sense that it can be adapted to any of the eight possible board geometries with just a local change [4] . It uses an attack table, for each side an array of 64 bytes, with following one- or two-bit attack counters per square ...

+-----+-----+-----+-----+-----+-----+-----+-----+
|   Pawns   |   Minors  |   Rooks   |Queen|King |
+-----+-----+-----+-----+-----+-----+-----+-----+
     7..6        5..4        3..2      1     0

... as used in move generation, SEE and evaluation.

Search

The search is a classical PVS iterative deepening approach with Zobrist key transposition table, quiescence search, null move pruning and mate distance pruning. Move ordering considers SEE and a simple killer heuristic.

Evaluation

Floyd's evaluation employs a vector of feature and weight pairs to calculate a score as weighted sum. In conjunction with a draw model [5] using sigmoid functions, the score is mapped to winning probabilities, suited for logistic regression tuning.

def evaluate(board, wiloVector, drawVector):
        wiloScore = ...snip... // a weighted sum of board features
        drawScore = ...snip... // another weighted sum of board features

        return sigmoid(drawScore) * 0.5
             + sigmoid(wiloScore)
             - sigmoid(wiloScore) * sigmoid(drawScore)

Misc

Floyd provides a Python API for search and evaluation functions [6] , i.e. for automated tuning [7] . It generates a compact KPK tablebase to deal with perfect knowledge, also available as stand alone project [8] [9] .

See also

Forum Posts

External Links

Chess Engine

Misc

References

Up one Level