Daniel R Schlegel and Stuart C Shapiro Department of Computer Science and Engineering University at Buffalo The State University of New York Buffalo New York USA lt drschlegshapiro gtbuffaloedu ID: 485226
Download Presentation The PPT/PDF document "Concurrent Reasoning with Inference Grap..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Concurrent Reasoning with Inference Graphs
Daniel R. Schlegel and Stuart C. Shapiro
Department of Computer Science and Engineering
University at Buffalo, The State University of New York
Buffalo, New York, USA<drschleg,shapiro>@buffalo.edu
D. R. Schlegel and S. C. Shapiro
1Slide2
Problem Statement
Rise of multi-core computersLack of concurrent natural deduction systemsApplication to natural language understanding
for terrorist plot detection.D. R. Schlegel2
A MotivationSlide3
What are Inference Graphs?
Graphs for natural deductionFour types of inference:ForwardBackwardBi-directional
FocusedRetain derived formulas for later re-use.Propagate disbelief.Built upon Propositional Graphs.Take advantage of multi-core computersConcurrency and schedulingNear-linear speedup.
D. R. Schlegel
3Slide4
Propositional Graphs
Directed acyclic graphEvery well-formed expression is a nodeIndividual constants
Functional termsAtomic formulasNon-atomic formulas (“rules”)Each node has an identifier, eitherSymbol, orwfti
[!]No two nodes with same identifier.D. R. Schlegel and S. C. Shapiro
4Slide5
Propositional Graphs
D. R. Schlegel and S. C. Shapiro5
If a, and b are true, then c is true.
a
c
and-ant
cq
wft1!
b
and-antSlide6
Inference Graphs
Extend Propositional GraphsAdd channels for information flow
(messages):i-channels report truth of an antecedent to a rule node.u-channels
report truth of a consequent from a rule node.Channels contain
valves.Hold messages back, or allow them through.D. R. Schlegel and S. C. Shapiro
6
i
-channel
u
-channel
a
c
and-ant
cq
wft1!
b
and-antSlide7
Messages
Five kindsI-INFER – “I’ve been inferred”U-INFER – “You’ve been inferred”
BACKWARD-INFER – “Open valves so I might be inferred”CANCEL-INFER – “Stop trying to infer me (close valves)”UNASSERT – “I’m no longer believed”D. R. Schlegel and S. C. Shapiro
7Slide8
Priorities
Messages have priorities.UNASSERT is top priorityCANCEL-INFER is nextI-INFER/U-INFER are higher priority closer to a result
BACKWARD-INFER is lowestD. R. Schlegel and S. C. Shapiro8Slide9
Rule Node Inference
Message arrives at node.
D. R. Schlegel and S. C. Shapiro9
a!
c
Assume we have a KB with a ^ b -> c, and b. Then a is asserted with forward inference.
A message is sent from a to wft1
and-ant
cq
a : true
wft1!
i
-infer
b
!
and-antSlide10
Rule Node Inference
2. Message is translated to
Rule Use InformationD. R. Schlegel and S. C. Shapiro
10
1 Positive Antecedent, a0 Negative Antecedents2 Total Antecedents
a : true
Rule Use Information is stored in rule nodes to be combined later with others that arrive.
a!
c
and-ant
cq
wft1!
b
!
and-antSlide11
Rule Node Inference
3
. Combine RUIs with any existing onesD. R. Schlegel and S. C. Shapiro
11
1 Positive Antecedent, b0 Negative Antecedents2 Total AntecedentsCombine the RUI for a with the one which already exists in wft1 for b.
1 Positive Antecedent,
a
0 Negative Antecedents
2
Total Antecedents
+
2
Positive Antecedents,
a,b
0 Negative Antecedents
2
Total Antecedents
=
a!
c
and-ant
cq
wft1!
b
!
and-antSlide12
Rule Node Inference
4. Determine if the rule can fire.
D. R. Schlegel and S. C. Shapiro12
We have two positive antecedents, and we need two. The rule can fire.
2 Positive Antecedents, a,b0 Negative Antecedents2
Total Antecedents
a!
c
and-ant
cq
wft1!
b
!
and-antSlide13
Rule Node Inference
5. Send out new messages.
D. R. Schlegel13
c will receive the message, and assert itself.
c : true
u-infer
a!
c
and-ant
cq
wft1!
b
!
and-antSlide14
Cycles
Graphs may contain cycles.No rule node will infer on the same message more than once.RUIs with no new information are ignored.
Already open valves can’t be opened again.D. R. Schlegel and S. C. Shapiro14
a
b
ant
cq
wft2!
wft1!
ant
cqSlide15
Concurrency and Scheduling
Inference Segment: the area between two valves.When messages reach a valve:A task
is created with the same priority as the message.Task: application of the segment’s function to the message.Task is added to a prioritized queue.Tasks have minimal shared state, easing concurrency.
D. R. Schlegel and S. C. Shapiro
15Slide16
Concurrency and Scheduling
A task only operates within a single segment.tasks for relaying newly derived information using segments “later” in the derivation are
executed before “earlier” ones, andonce a node is known to be true or false, all tasks attempting to derive it are canceled, as long as their results are not needed elsewhere.
D. R. Schlegel and S. C. Shapiro16Slide17
Example
D. R. Schlegel and S. C. Shapiro17
cq
Backchain
on
cq
. Assume every node requires a single one of its incoming nodes to be true for it to be true
(simplified for easy viewing). Two processors will be used.Slide18
Example
D. R. Schlegel and S. C. Shapiro18
Backward Inference
(Open valve)
Inferring
Inferred
Cancelled
cqSlide19
Example
D. R. Schlegel and S. C. Shapiro19
Backward Inference
(Open valve)
Inferring
Inferred
Cancelled
cqSlide20
Example
D. R. Schlegel and S. C. Shapiro20
Backward Inference
(Open valve)
Inferring
Inferred
Cancelled
cqSlide21
Example
D. R. Schlegel and S. C. Shapiro21
Backward Inference
(Open valve)
Inferring
Inferred
Cancelled
cqSlide22
Example
D. R. Schlegel and S. C. Shapiro22
Backward Inference
(Open valve)
Inferring
Inferred
Cancelled
cqSlide23
Example
D. R. Schlegel and S. C. Shapiro23
Backward Inference
(Open valve)
Inferring
Inferred
Cancelled
cqSlide24
Example
D. R. Schlegel and S. C. Shapiro24
Backward Inference
(Open valve)
Inferring
Inferred
Cancelled
cqSlide25
Example
D. R. Schlegel and S. C. Shapiro25
Backward Inference
(Open valve)
Inferring
Inferred
Cancelled
cqSlide26
Example
D. R. Schlegel and S. C. Shapiro26
Backward Inference
(Open valve)
Inferring
Inferred
Cancelled
cqSlide27
Example
D. R. Schlegel and S. C. Shapiro27
Backward Inference
(Open valve)
Inferring
Inferred
Cancelled
cqSlide28
Example
D. R. Schlegel and S. C. Shapiro28
Backward Inference
(Open valve)
Inferring
Inferred
Cancelled
cqSlide29
Example
D. R. Schlegel and S. C. Shapiro29
Backward Inference
(Open valve)
Inferring
Inferred
Cancelled
cqSlide30
Example
D. R. Schlegel and S. C. Shapiro30
Backward Inference
(Open valve)
Inferring
Inferred
Cancelled
cqSlide31
Example
D. R. Schlegel and S. C. Shapiro31
Backward Inference
(Open valve)
Inferring
Inferred
Cancelled
cqSlide32
Example
D. R. Schlegel and S. C. Shapiro32
Backward Inference
(Open valve)
Inferring
Inferred
Cancelled
cqSlide33
Example
D. R. Schlegel and S. C. Shapiro
33
Backward Inference
(Open valve)
Inferring
Inferred
Cancelled
cqSlide34
Example
D. R. Schlegel and S. C. Shapiro
34
Backward Inference
(Open valve)
Inferring
Inferred
Cancelled
cqSlide35
Example
D. R. Schlegel and S. C. Shapiro35
Backward Inference
(Open valve)
Inferring
Inferred
Cancelled
cqSlide36
Evaluation
Concurrency:Near linear performance improvement with the number of processorsPerformance
robust to graph depth and branching factor changes.Scheduling Heuristics:Backward-inference with or-entailment shows 10x improvement over LIFO queues, and 20-40x over FIFO queues.
D. R. Schlegel and S. C. Shapiro
36Slide37
Acknowledgements
This work has been supported by a Multidisciplinary University Research Initiative (MURI) grant (Number W911NF-09- 1-0392) for Unified Research on Network-based Hard/Soft Information Fusion, issued by the US Army Research Office (ARO) under the program management of Dr. John Lavery.
D. R. Schlegel37