Advanced Artificial Intelligence Lecture 22: Wong

1
Advanced Artificial IntelligenceLecture 22
Wong LeungEvolutionary Program Induction
directed by Logic Grammars

• Bob McKay
• School of Computer Science and Engineering
• College of Engineering
• Seoul National University

2
Outline

• Logenpro
• Definite Clause Grammars
• Grammars for type restrictions
• Grammars to incorporate syntactic domain
  knowledge
• Grammars to incorporate semantic domain knowledge?

3
History

• 1992 Cohen, CFG to control ILP search space
• 1994/95 - grammars to control GP search space
• Whigham
• Wong Leung
• Geyer-Schulz

4
Why grammars?

• Language bias
• Search bias
• Type constraints
• Homologous crossover

5
Leung Kwong Sak Wong Man Leung

• Originally Chinese University of Hong Kong
• Wong now at Lingnan University
• Very strong group in evolutionary computation and
  genetic programming
• Logenpro GP system 1995, led to book in late
  1990s

6
Logenpro

• Grammar-based genetic programming system
• Defines a genetic programming system constrained
  by a grammar
• Like Whigham, evolves derivation trees
• Main difference definite clause logic grammars
• A significant extension of context free grammars
• Secondary difference Lisp
• Rather than evolving expressions, logenpro
  evolves complete Lisp programs

7
Definite Clause Grammars

• Context free grammars extended in two ways
• Any symbol (termnal or non-terminal) may have
  arguments
• Generally logical variables
• Additional (logic) clauses may impose semantic
  constraints on variables

8
Definite Clause Grammar Example

• a-useful-program - first-component(?X), gramm
  ar rule with parameter is-useful(?X, ?Y),
  semantic constraint on parameters second-comp
  onent(?Y). additional grammar rule with
  parameter

9
Boolean Grammar Example

• start -gt (defun parity (L) ,s-expr(boolean),
  )) . s-expr(BOOLEAN) -gt T .
  s-expr(BOOLEAN) -gt nil . s-expr(BOOLEAN) -gt
  ( , op, s-expr(BOOLEAN), s-expr(BOOLEAN),
  ) . s-expr(BOOLEAN) -gt ( , parity
  ,s-expr(LIST), ) . s-expr(BOOLEAN) -gt ( ,
  first ,s-expr(LIST), ) . s-expr(LIST) -gt
  L . s-expr(LIST) -gt s-expr(SMALLER-LIST).s-
  expr(SMALLER-LIST)-gt ( , rest ,
  s-expr(LIST),). op -gt AND . op -gt OR
  . op -gt NAND . op -gt NOR .

10
Adding Restrictions

• start -gt (defun parity (L) , (ifnil L T
  ,s-expr(boolean), )) . s-expr(BOOLEAN) -gt
  T . s-expr(BOOLEAN) -gt nil .
  s-expr(BOOLEAN) -gt ( , op, s-expr(BOOLEAN),
  s-expr(BOOLEAN), ) . s-expr(BOOLEAN) -gt (
  , parity , s-expr(SMALLER-LIST), ) .
  s-expr(BOOLEAN) -gt ( , first
  ,s-expr(LIST), ) . s-expr(LIST) -gt L .
  s-expr(LIST) -gt s-expr(SMALLER-LIST).s-expr(SMA
  LLER-LIST)-gt ( , rest , s-expr(LIST),).
  op -gt AND . op -gt OR . op -gt NAND
  . op -gt NOR .

11
Experiments in Paper

• Even-n-parity for fixed n
• Even-n-parity general recursive function
• using highly constrained grammar to limit the
  search space

12
Definite Clause vs Context Free

• In theory, definite clause grammars significantly
  extend context free grammars
• Is this important?
• Almost all applications only use the context-free
  sub-part of definite clause grammars
• In particular, the experiments in the paper do
  not use the extensions in any important way
• Everything can be done with context free grammars
• Is it worth the extra computational cost?

13
Summary

• Logenpro
• Definite Clause Grammars
• Grammars for type restrictions
• Grammars to incorporate syntactic domain
  knowledge
• Grammars to incorporate semantic domain knowledge?

14
?????