From Chessprogramming wiki
Jump to: navigation, search

Home * Engines * Octavius

a chess engine by Luke Pellen relying purely on an evaluation based on a four-layer neural network without any material analysis, choosing the move with the best score on its single output neuron, performing a one ply look ahead. Despite Sebastian Thrun's remark, that using the raw board representation as input of a neural network is a poor choice [2], but quite similar to one input representation of the Neural MoveMap Heuristic by Levente Kocsis et al. [3], Octavius has an input layer of 768 (12x64) nodes, a sequence of 12 numbers (11 or 12 times 0.0, 1 or 0 times 1.0) for each square, where a 1.0 was given for the particular piece (if any) residing on that square. Two hidden layers with 1024 and 512 nodes are feed forward to one output node, which represents the score of the position for one side to move. In total, Octavius' net has 2305 nodes with 1,311,232 connections.

Selected Games

Beginning in 1999, Octavius performed 284,552 backpropagations up to April 26, 2004, when it won its first game against his creator [4]:

[Event "Octavius Test"]
[Site "Lukes Computer"]
[Date "2004.04.26"]
[Round "?"]
[White "Luke Pellen"]
[Black "Octavius P4Oct4.ann"]
[Result "0-1"]

1.e4 Nf6 2.Nc3 e5 3.Bb5 c6 4.Ba4 Bc5 5.Nf3 O-O 6.O-O d6 7.d3 Bg4 
8.Bb3 Nh5 9.a3 Re8 10.Ba2 Qf6 11.Bg5 Qg6 12.b4 Bb6 13.Bh4 Nf4 
14.Kh1 Nd7 15.g3 Nh3 16.Qe2 h5 17.Rae1 d5 18.exd5 cxd5 19.Nxd5 e4 
20.dxe4 Rxe4 21.Qxe4 f5 22.Qe7 Bxf3# 0-1

See also


Forum Posts

External Links


Up one Level