ICL Slides

advertisement
The Independent Choice Logic
for modeling multiple agents
under uncertainty
David Poole
Http://www.cs.ubc.ca/spider/poole
Presented by Mei Huang
Mar. 18,2005
Outline
Knowledge Representation
ICL Overview
ICL Semantics
ICL Influence
Discussion
Outline
Knowledge Representation
ICL Overview
ICL Semantics
ICL Influence
Discussion
“a bridge between logical representation and belief networks”
Q
?
Discussion
What do you think of the author’s claim
that using disjunction to handle
uncertainty is a stupid thing?
From the view point of representation,
disjunction is not that bad, at least, it
provides compact representations than
ICL. E.g, “ea1v …v an” in ICL needs n
expressions.
Outline
Knowledge Representation
ICL Overview
ICL Semantics
ICL Influence
Discussion
Note: each item in F, i.e, each rules has an associated probability. E.g, for
the first rule, the probability is P(f | c1, b1). This probability is different from
the probability of a possible world.
Q
?
Discussion
 How to cast the student-take-course domain
problem into ICL?
 Choice Space C={{L1,L2,L3},{High,Normal,Easy}}
Facts F={Grade=A  IQ=L3 ^ Difficulty=Easy;
Grade=B  IQ=L2 ^ Difficulty=High;
……}
Give P0(L1), P0(L2), P0(L3), P0(High), P0(Normal),
P0(Easy)
Outline
Knowledge Representation
ICL Overview
ICL Semantics
ICL Influence
Discussion
Q?
Discussion
Why it looks so straight forward to map a
decision tree to the rules of ICL?
The logical part of ICL focuses on
attributes’ values, not attributes. ICL’s Fact
set F interprets CPDs of BLP or LoPRM.
Q?
Discussion
 How to translate this BN into ICL?
Burglary
earthQuake
Alarm
 It is better to redraw the graph as 8 belief networks as
follows:
B=true
Q=true
A=true
……
Because ICL’s rules are defined on the value level.
are mutually exclusive
Outline
Knowledge Representation
ICL Overview
ICL Semantics
ICL Influence
Discussion
Discussion
 In the diagram in slide15, why the arrows are
uni-directional? Can they be made bidirectional?
 What can ICL buy us? I.e, is this explicit bridge
between logic and decision theory necessary or
is the gap between the two enlarged imaginarily
by the author? (when we deal with decision
projects, logic is usually embodied implicitly in
the program)
Download