Transcript BLOG model

Topics
• Combining probability and first-order logic
– BLOG and DBLOG
• Learning very complex behaviors
– ALisp: hierarchical RL with partial programs
• State estimation for neurotrauma patients
– Joint w/ Geoff Manley (UCSF), Intel, Omron
•
•
•
•
Metareasoning and bounded optimality
Transfer learning (+ Jordan, Bartlett, MIT, SU, OSU)
Knowing everything on the web
Human-level AI
Heart rate
Blood pressure
Oxygen saturation
Sedation level
Cardiac output
ICP wave
Pulmonary artery pressure
Intracranial pressure
Temperature
End expiratory pressure
Respiratory rate
Tissue oxygen
Ventilation mode
Inspired oxygen
Tidal volume
Peak pressure
State estimation: 3x5 index card
Patient 2-13
Dynamic Bayesian Networks
DBNs contd:
Research plan
• DBN model: ~200 core state variables, ~500
sensor-related variables
• Learn model parameter distributions from DB
• Infer patient-specific parameters online
• Goals:
–
–
–
–
Improved alarms
Diagnostic state estimation => improved treatment
Solve the treatment POMDP
Structure discovery => better understanding of
physiology
Possible worlds
• Propositional
• First-order + unique names, domain closure
A
B
A
B
A
B
A
B
A
B
C
D
C
D
C
D
C
D
C
D
• First-order open-world
ABCD
ABCD
ABCD
ABCD
ABCD
ABCD
Example: Citation Matching
[Lashkari et al 94] Collaborative Interface Agents, Yezdi
Lashkari, Max Metral, and Pattie Maes, Proceedings of
the Twelfth National Conference on Articial
Intelligence, MIT Press, Cambridge, MA, 1994.
Metral M. Lashkari, Y. and P. Maes. Collaborative
interface agents. In Conference of the American
Association for Artificial Intelligence, Seattle, WA,
August 1994.
Are these descriptions of the same object?
What authors and papers actually exist, with what
attributes? Who wrote which papers?
General problem: raw data -> relational KB
Other examples: multitarget tracking, vision, NLP
Approach: formal language for specifying first-order
open-world probability models
BLOG generative process
• Number statements describe steps that add
some objects to the world
• Dependency statements describe steps that
set the value of a function or relation on a
tuple of arguments
• Includes setting the referent of a constant
symbol (0-ary function)
• Both types may condition on existence and
properties of previously added objects
BLOG model (simplified)
guaranteed Citation Cit1, Cit2, Cit3, Cit4, Cit5, Cit6, Cit7;
#Researcher ~ NumResearchersPrior();
Name(r) ~ NamePrior();
#Paper(FirstAuthor = r) ~ NumPapersPrior(Position(r));
Title(p) ~ TitlePrior();
PubCited(c) ~ Uniform({Paper p});
Text(c) ~ NoisyCitationGrammar
(Name(FirstAuthor(PubCited(c))), Title(PubCited(c)));
Basic results
• Theorem 1: Every well-formed* BLOG model
specifies a unique distribution over possible
worlds
• The probability of each (finite) world is given
by a product of the relevant conditional
probabilities from the model
• Theorem 2: For any well-formed BLOG
model, there are algorithms (LW, MCMC) that
converge to correct probability for any query,
using finite time per sampling step
Citation Matching Results
Error
(Fraction of Clusters Not Recovered Correctly)
0.25
0.2
Phrase Matching
[Lawrence et al. 1999]
0.15
Generative Model + MCMC
[Pasula et al. 2002]
Conditional Random Field
[Wellner et al. 2004]
0.1
0.05
0
Reinforce
Face
Reason
Constraint
Four data sets of ~300-500 citations, referring to ~150-300 papers
DBLOG
• BLOG allows for temporal models – time is just a
logical variable over an infinite set
• Inference works (only finitely many relevant
random variables) but is grossly inefficient
• DBLOG includes time as a distinguished type and
predecessor as distinguished function; implements
special-purpose inference:
– Particle filter for temporally varying relations
– Decayed MCMC for atemporal relations
Open Problems
• Inference
– Applying “lifted” inference to BLOG (like Prolog)
– Approximation algorithms for problems with huge/growing
numbers of objects
• Knowledge representation
– Hierarchical activity models
– Undirected submodels
– Nonparametric extensions (cf. de Freitas, 2005)