Selmer Bringsjord & Kostas Arkoudas Andy Shilliday, Josh Taylor, Sunny Khemlani Rensselaer AI &...
-
Upload
lionel-small -
Category
Documents
-
view
217 -
download
1
Transcript of Selmer Bringsjord & Kostas Arkoudas Andy Shilliday, Josh Taylor, Sunny Khemlani Rensselaer AI &...
Selmer Bringsjord & Kostas ArkoudasAndy Shilliday, Josh Taylor, Sunny Khemlani
Rensselaer AI & Reasoning (RAIR) LabDepartment of Cognitive ScienceDepartment of Computer Science
{selmer,arkouk,shilla,tayloj,khemls}@rpi.edu
DARPA IPTO 3.31.05
Poised-For Learning:Update on RAIR Lab’s Progress as of March 31 2005
(project start date: Oct 1 2004)
PFL(Overview Figure)
The Six Distinguishing Attributes of Poised-For Knowledge
•Attribute 1: Mixed Representation Mode
• Symbolic and Diagrammatic
•Attribute 2: Tapestried
•Attribute 3: Extreme Expressivity
•Attribute 4: Mixed Inference Types
•Attribute 5: Deep Connection to Natural Language
•Attribute 6: Multi-Agent Structures{ }
{ }QuickTime™ and a
Graphics decompressorare needed to see this picture.
{ }“The Eye”demo
What Systems Can Handle This?
MARMMLAttributes 1-3 Attributes 4-6
... + Athena + ...
... Vampire ... ... Paradox ...
Guaranteed soundness
And Athena, in 2005?
•Used for dataflow analysis
•Used for verification of INS
•UNIX-style file system formal analysis
•Generic software verification
•Code-carrying proofs
•Cryptographic applications
•And of course: Athena is one of the cornerstones of the revolutionary approach to learning by reading known as poised-for learning.
Now, to the SOW, workplan, and status...
PFL (SOW-annotated)
• Stage I: Formally specify and implement the formal visual/mental models scheme for handling mixed-mode representation (Attribute 1) that uses not only code in Athena's underlying denotational proof language (DPL), but also enhancements of “Barwisean/Johnson-Lairdian” diagrams discussed in the proposal proper.
• Stage II: Hand-construct p.f.-knowledge $\Pi^h_1, \Pi^h about domain D that has various permutations of the six attributes. This p.f.-knowledge will be built from elements prior to the arrival of `$ in the overview figure.
• Stage III: Implement an Athena/MARMML-based system that automatically generates, from $\Pi^h_1, \Pi^h., the representation of an answer paired with an accompanying justification rep( A, J) in response to queries about the domain D.
Stages I-III
• Stage IV: Implement an Athena/MARMML-based system that automatically generates, from the representation of an answer and accompanying justification rep(A, J) in Stage III, the corresponding output O in English.
• Stage V: Implement an Athena/MARMML-based system that automatically p.f.-learns $\Pi^a_1, \Pi^a about domain D from post-NLU content (input rep(I) from reading material, internal anticipatory queries Q
1, Q
2, ... Q
n, and prior knowledge $\Ps .
Stages IV & V
•Stage I: Formally specify and implement the formal visual/mental models scheme for handling mixed-mode representation (Attribute 1) that uses not only code in Athena's underlying denotational proof language (DPL), but also enhancements of “Barwisean/Johnson-Lairdian” diagrams discussed in the proposal proper.
On Stage I...
Facts Re. Diagrammatic Learning & KR&R
• The most powerful cognitive systems represent knowledge, and reason over that knowledge, in irreducibly visual/diagrammatic fashion.
• For confirmation one can consult a good cognitive psychology text, e.g., Goldstein’s Cognitive Psychology.
• These cognitive systems learn in in large part by reading content that, in turn, is in large part diagrammatic.
• In some of the texts in our library for the project, more space is devoted to pictographic content than textual content.
• When it comes to reasoning in support of learning by reading, we now know that there is overwhelming empirical evidence that humans reason is both “proof-theortic” and “mental models-based” fashion (Johnson-Laird, Rips, Bringsjord & Yang).
Prior AI-Relevant Work
•Diagrammatic processing irreducibly visual in Hyperproof (Barwise, Etchemendy, Barker-Plummer)
•“Fake” diagrammatic reasoning in IDR” (et al)
•Rigorous recasting of diagrams in logicist fashion (Wang, Lee, & Zeevat; Barker-Plummer’s GROVER)
The Dream
Blocks World Module
Digraphic Module
Venn Diagram Module
?Line & Angle Module
Engineering Reality
Blocks World Module
Digraphic Module
Venn Diagram Module
?Line & Angle Module
The ‘New Order’ Scenario
•John H. was killed by a member of the Al-Qaeda cell 'The New Order'.
•The only members of 'The New Order' were John H., Majed H., and Essid D.
•Within-cell killings only occur when the attacker believes the victim is a traitor, and never when the attacker is of lower rank.
•Essid D. believes that nobody is a traitor who John H. believes is a traitor.
•John H. believes everyone except Majed H. is a traitor.
•Majed H. believes that everyone who is not of lower rank than John H. is a traitor.
•Majed H. believes everyone is a traitor who John H. believes is a traitor.
•No one believes everyone in 'The New Order' is a traitor.
P1
P2
P3P4
P7P6P5
Paradox generates this:
Which is translated to
MDF
Which is translated to
MDF
Neither seems quite as nice as
this digraph:
Prior AI-Relevant Work
•Diagrammatic processing irreducibly visual in Hyperproof (Barwise, Etchemendy, Barker-Plummer)
•“Fake” diagrammatic reasoning in IDR” (et al)
•Rigorous recasting of diagrams in logicist fashion (Wang, Lee, & Zeevat; Barker-Plummer’s GROVER)
Graphical Signatures etc. nice fit with Athena
• Stage II: Hand-construct p.f.-knowledge $\Pi^h_1, \Pi^h about domain D that has various permutations of the six attributes. This p.f.-knowledge will be built from elements prior to the arrival of `$ in the overview figure.
• Stage III: Implement an Athena/MARMML-based system that automatically generates, from $\Pi^h_1, \Pi^h., the representation of an answer paired with an accompanying justification rep( A, J) in response to queries about the domain D.
Stages II & III
Math Example #5 (”Parallel Lines”)
(Gr 7 Textbook)
Query Q(TIMSS M8 2003)
Q1
Q2
O = (J, A)
Math Example #5 (”Parallel Lines”) Query Q
(TIMSS M8 2003)
O = (J, A)
Astronomy Example #1 (”Solar System”)
Query Q
O = (J, A)
Is every planet inside the asteroid belt smaller than the sun?
Astronomy Example #1 (”Solar System”)
Query Q
O = (J, A)
Is every planet inside the asteroid belt smaller than the sun?
•Stage IV: Implement an Athena/MARMML-based system that automatically generates, from the representation of an answer and accompanying justification rep(A, J) in Stage III, the corresponding output O in English.
Stage IV...
Prior R&DPROVERB But...
Taps into “unprincipled” NLG
No natural langugage corresponding to diagrammatic knowledgeCan’t handle resolution-based reasoning
Can’t handle methods, only proofs (not dynamic proofs)
Dormant?
Reasoning that is input lacks power of Athena
First, NDL Proofs to English
Current Status•Project right on schedule and proceeding
according to plan; the SOW blueprint being followed, and proving to be a clean, productive workplan.
•P-f Learning a new form of learning at the heart of learning by reading.
•Diagrammatic KR&R in support of learning by reading promises to be a substantive advance for AI.
•Now have some implemented p-f knowledge, related to texts in our domains, and vision in original white paper turning into concrete reality.
•Good progress on the automatic production of natural language answer and justification, showing that learning has taken place.
•Core algorithms for Stage V: Next update!
But how??
program synthesis; plan/method generation
The End (of DARPA content;
remaining slides just content to possibly pull from)
w/i team notes...
Math Example #7 (”Induction”)
(Gr 7 Textbook)
Query Q(TIMSS M8 2003)
O = (J, A)Q1
Q2
Astronomy Example #2 (”Epistemic”)
----------
Query Q?
O = (J, A)Q1
Q2
Won’t Discuss, Here, Connections to Slate; But Later?
Text
QuickTime™ and aGraphics decompressor
are needed to see this picture.
Bringsjord, Shilliday, Taylor, Arkoudas, Khemlani
Overall Objective“The objective of this effort is to investigate new architectures, algorithms, and designs to lead to: implemented machine reasoning over knowledge in expressive formats that include doxastic (= epistemic) and deontic operators; understanding of the tractability of using such implementations in a multi-agent setting; and transfer of such architectures, algorithms, and eventual implementations into other relevant DoD-related efforts.”
Epistemic Work...“Rendering Scenarios Expressed Doxastic Systems (e.g., KD45) in Computational Form via Logic-Based AI Techniques. The contractor shall develop the theoretical constructions (architectures, algorithms, designs, etc.) necessary for the computational implementation, in the contractor’s systems, of test scenarios expressed in the modal logic KD45 (and/or other such logics) of belief and knowledge. This logic allows for reasoning over doxastic information, which is information about what agents believe and know. The contractor shall provide a method to address the technical problem known as ‘logical omniscience.’”
Militaristic “Wise Man”** A L E R T **
To: Special Forces Company DFrom: Central Command, Integrated Special Forcescc: Special Forces Companies A, B, C
Recent HUMINT and SIGINT reveals that at least one of you (A, B, C, D), at present, has been locked in as a target of MET's highly effective medium-range laser-guided missile system, the Azan+. Despite the threat this poses (launch could come at any moment), under no circumstances should you change your present location: Any movement could result in your being locked into the sites of the Azan+, if you aren't already. The last thing we want is for a group that isn't locked in to be successfully targeted.
As you know, and as the other companies know as well, you cannot determine through use of your EYE system whether your own company has been locked in by the Azan+'s targeting system. But the EYE *can* determine whether *another* company has been locked in (a signature laser tag is visible to the EYE when the Azan+ is aimed at units other than yours). All of you, as you know, can scan each other with the EYE. ...
Militaristic “WM” (con)
Company A, upon receiving an alert a few minutes ago informing it that at least one of A, B, C, and D is locked in, and asking it to respond as to whether or not it can infer that it is locked in, engaged its EYE and then sent out comm declaring that it does not know whether it is locked in. After this same comm, B issued the same message, and then C received the same comm and soon thereafter radioed the same message. Now the ball is in your court.
As you know, if a company is currently locked inby the Azan+, certain jamming techniques implemented from our location can cloak you once again -- but if these jamming techniques are used mistakenly, if they are used when you are *not* already locked in by the Azan+, you will be immediately targeted, and launch will almost certainly ensue shortly thereafter.
We await your response.
Wise Man Puzzle
Athena Demo
QuickTime™ and aAnimation decompressor
are needed to see this picture.
Deontic Work...
“Rendering Scenarios Expressed Deontic Systems (e.g., DSDL3) in Computational Form via Logic-Based AI Techniques. The contractor shall develop the theoretical constructions necessary for the computational implementation, in the contractor’s systems, of test scenarios expressed in the modal logic DSDL3 (Lewis 1974) (and/or other such logics) of obligation. The contractor shall address the technical problem known as ‘adequacy’ (or conditional obligation).”
Empirical Investigation...
•Is reasoning about norms a verifiable human activity?
•Does deontic logic provide an adequate representational system within which to express this activity?
The Human Case
Yes
Yes
(Bello & Yang forthcoming)
Oxy vs Jihad
Focussing In (e.g.)
...
In Deontic Logic(glimpse ahead)
•Delivery of expanded Springer-Verlag paper, with acknowledgement to AFRL-Rome.
•Finish deontic case as we did for the epistemic case:
•Mechanize one or more expressive deontic logic in Athena.
•Use that mechanization to solve queries concerning situations in “Oxy v Jihad” game
•Submit to AAAI Fall Symposium on Machine Ethics.
Next Steps
TowardAthena/MARMML/SNeRE-
Powered Intelligent Agents...
Overview of Logic-Based Agent
(review; Nilsson)
QuickTime™ and aGraphics decompressor
are needed to see this picture.
Overview of Logic-Based Agent
(AIMA2e)
Survivability/Extensibility
3rd Gen Systems RAIR L Agents
Synthetic Characters...
(for Wargaming)
FSA-Level Characters
Advanced Synthetic Characters
E.
Bringsjord,McEvoy,
Destefano
Building a Virtual Person (E)
from the ‘Dark Side’”
RASCALS
RASCALS logic-based
SNeRESNeRE
AthenaAthena
•Napolean at Waterloo
•Patrol
•...
Electrifying Wargames?
The End