Knowledge Acquisition from Game Records

Knowledge Acquisition
from Game Records
Takuya Kojima, Atsushi Yoshikawa
Dept. of Computer Science and Information Engineering
National Dong Hwa University
Reporter:Lo Jung-Yun
Outline
•
•
•
•
Introduction
A Deductive Approach
An Evolutionary Approach
Conclusions
2
Introduction
3
Purpose
• The knowledge of human experts has
two important features: quality and
quantity
• Some systems have tried to acquire
Go knowledge, most of them acquire
only fixed-shaped knowledge
• A new algorithm which yields more
flexible knowledge is therefore
necessary
4
Classification of Go knowledge
• Classify Go knowledge according to
two criteria
– Form
• Patterns
• Sequence of moves
• maxims
– Degree of validity
• Strict knowledge
• Heuristic knowledge
5
Two Approaches
• This paper focuses on pattern
knowledge
Deductive Approach
Strict
Knowledge
Several rules are acquired
from a single training example
Evolutionary Approach
Heuristic
Knowledge
Acquire a large amount of
heuristic knowledge from a large
amount of training examples
6
A Deductive Approach
7
System overview
8
Model introduction
•Knowledge base
–Basic rules
–Forcing rules
rulenfor : if (cond1  ...  condm )then( ( BC*, s)( s  x, y, t ))
•Decision maker
9
Rule acquisition algorithm
Chooses good moves to
be learned
Extracts relevant parts
from board configuration
Generalizes the position
and the move
10
An Evolutionary Approach
11
Concept
• Each rule takes the form of a
production rule
• There are no rules in the initial state
• Feed, consume, and split
– with activation value
12
Algorithm
13
Rules
• Feeding
– When five rules are matched…
• Consuming
– Each rule consumes activation value at each
step
– Rule whose activation value is 0 die
• Splitting
– If activation value is greater than threshold –
split it!
• Original rules → “parent”
• Randomly add a new condition from among the
objects on the current board
14
Application to Tsume-Go
• Maybe many rules apply in the same
situation
– Assign priority
• Priority assignment algorithm
– Assignment of weight to rules
– Probability of rule accuracy
15
Application to Tsume-Go
• Compare with two algorithm
– Fixed algorithm
– Semi-fixed algorithm
16
Application to Tsume-Go
17
Conclusions
• Explain 2 approaches:
– Deductive
– Evolutionary
• The performance is as good as 1 dan
human players
18