slides of Presentation - VUB Parallel Computing Laboratory

An Alternative Approach for
Playing Complex Games like
Chess.
Jan Lemeire
May 19th 2008
Alternative Game Playing Approach
1
Research
Topics
Computer
versus Brain
Deep Blue: 600 million
evaluations/second
Alternative Game Playing Approach
Jan Lemeire
Chess experts: 10
patterns/second
Pag. 2 / 14
Brute force chess playing
White plays
Most succesful!
Black plays
White plays
Black plays
Evaluation of future states
Alternative Game Playing Approach
Jan Lemeire
Pag. 3 / 14
Alternative game playing
approaches
Decision-making: used to map states
to operators.
Explanation-Based Learning (EBL): try
to learn the states that lead to
advantageous situations.
States are identified by patterns.
Alternative Game Playing Approach
Jan Lemeire
Pag. 4 / 14
Game State Evaluation
All approaches rely on a game state
evaluation:
– measure ‘goodness’ of state
Or
– to select a good move.
My hypothesis is based on a problem with state
evaluation.
Alternative Game Playing Approach
Jan Lemeire
Pag. 5 / 14
Example: fork pattern.
Opportunity!
Fork pattern is way to
success!
Alternative Game Playing Approach
Jan Lemeire
Counter move…
Fork pattern does not
give an advantage…
Pag. 6 / 14
Correct evaluation problem
Consider two relevant patterns, P1 and P2
 evaluation = f(P1, P2)
 4 regions in state space should be considered:
P1
P1
P1
P1
&
&
&
&
P2
P2
P2
P2
For example ‘fork’ and ‘make chess’
Alternative Game Playing Approach
Jan Lemeire
W: advantageous for white
B: advantageous for black
0: no advantage
Pag. 7 / 14
Correct evaluation problem
Evaluation = f(features or patterns).
One has to capture all situations in which the
pattern leads to a successful outcome, all
counter plans have to be excluded.
Evaluation of pattern combinations heavily
depends on game context!
Features alone do not give us the right
information.
Alternative Game Playing Approach
Jan Lemeire
Pag. 8 / 14
Known problem in literature
Deep Blue relied on looking as far as possible into the
future and just a simple state evaluation.
“However, even simple patterns like a knight fork are
non-trivial to formalize…” Fürnkranz 2001.
“Learning too many too specialized rules with
explanation-based learning”, Minton 1984, Epstein,
Gelfand and Joanna Lesniak 1996 (HOYLE pattern-based
learning).
“Even in simple games, such as tic-tac-toe, 45 concepts
were learned with 52 exception clauses”, Fawcett and
Utgoff 1991.
Jan Lemeire
Pag. 9 / 14
Hypothesis
“The impact of a pattern on the outcome
of the game entirely depends on
whether or not some states, called the
effects, are attained during the
continuation of the game.”
Alternative Game Playing Approach
Jan Lemeire
Pag. 10 / 14
New kind of knowledge
Fork  win a piece
Weak king’s defense  successful
attack on the king
Pressure  successful combination
Patterns denote opportunities,
advantages have to be verified.
Alternative Game Playing Approach
Jan Lemeire
Pag. 11 / 14
Alternative game playing
White recognizes
pattern 1: 1
White has to check in
game tree whether
- a positive effect can
be attained: 1
- black can neutralize
pattern 1: 1
Hypothesis:
More efficient than brute
force tree exploration
Alternative Game Playing Approach
Jan Lemeire
Pag. 12 / 14
Similar to human game playing!
Chess experts rely on falsification (Cowley and
Byrne, 2004).
Humans can easily recognize and identify
patterns, but have difficulties formally defining
them.
Humans can pinpoint the patterns that were
decisive in a game, can answer why-questions.
not by current computer game playing
Humans can reason about a game.
Alternative Game Playing Approach
Jan Lemeire
Pag. 13 / 14
Hypothesis requires theoretical or
experimental confirmation…
Test
by simulation of games  no decisive conclusion yet.
Pattern engine needed that is able to:
Describe patterns
Recognize patterns
Extract patterns
Reason with patterns
“White attacks two black pieces with a fork, one of the pieces
can make chess. White thus has to move its king and black
can bring his second piece into safety.”
Theoretical proof?
Alternative Game Playing Approach
Jan Lemeire
Pag. 14 / 14