Nate Derbinsky University of Michigan Why Declarative LongTerm Memory Systems Guiding Principle Explore and evaluate new architectural mechanisms if a gents are missing important functionality ID: 919750
Download Presentation The PPT/PDF document "Episodic & Semantic Soar’s Long-Te..." 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
Episodic & SemanticSoar’s Long-Term Declarative Memory Systems
Nate Derbinsky
University of Michigan
Slide2Why Declarative Long-Term Memory Systems?
Guiding Principle
. Explore and evaluate new architectural mechanisms
if…agents are missing important functionality andexisting mechanisms cannot effectively and efficiently support this functionality
9 June 2011
Soar's Long-Term Declarative Memory Systems
2
Slide3Soar 9
9 June 2011
Soar's Long-Term Declarative Memory Systems
3
Slide4Autobiographic Agent HistoryWhat
The ability to remember, in context, experience that wasn’t necessarily known to be important
Why
Virtual Sensing [Nuxoll & Laird 2007]Action Modeling [
Xu & Laird 2010]…
9 June 2011
Soar's Long-Term Declarative Memory Systems
4
Arguments for a distinct mechanism
Rule matching scales with working memory size
Rule-based episodes would be problematic to encode and retrieve
Slide5Episodic Memory: Big Picture
Episodic memory is a
weak
learning mechanismAutomatically captures, stores, and temporally indexes agent stateSupports content-addressable agent interface to autobiographical prior experience9 June 2011Soar's Long-Term Declarative Memory Systems
5
Slide6Architectural Integration
9 June 2011
6
Soar's Long-Term Declarative Memory Systems
Slide7Architectural Integration
9 June 2011
7
Soar's Long-Term Declarative Memory Systems
Slide8Architectural Integration
9 June 2011
8
Soar's Long-Term Declarative Memory Systems
Slide9Architectural Integration
9 June 2011
9
Soar's Long-Term Declarative Memory Systems
Slide10Mechanism Detail
Encoding
Faithful “snapshot” of top-state
StorageNo dynamics (such as forgetting)RetrievalNearest-neighbor using structural cue matchingBiased by recencyPartial match
9 June 2011
Soar's Long-Term Declarative Memory Systems
10
Slide11Select Prior Work
Architectural Integration & Agent Capabilities
Extending Cognitive Architecture with Episodic Memory
Andrew Nuxoll & John Laird. AAAI (2007)EfficiencyEfficiently Implementing Episodic Memory
Nate Derbinsky & John Laird. ICCBR (2009)Integrating Learning Mechanisms
Learning to Use Episodic Memory
Nicholas
Gorski
& John Laird. Cognitive Systems Research (2011)
Using Diverse Cognitive Mechanisms for Action Modeling
John Laird, Joseph
Xu
, & Sam
Wintermute. ICCM (2010)
9 June 2011
Soar's Long-Term Declarative Memory Systems11
Slide12Soar 9
9 June 2011
Soar's Long-Term Declarative Memory Systems
12
Slide13Declarative Knowledge StoreWhat
The ability to store and flexibly retrieve large amounts of knowledge about the world, independent of the context in which it was originally learned
Why
Large, diverse KBsLexical (WordNet
)Ontological (Cyc)
…
9 June 2011
Soar's Long-Term Declarative Memory Systems
13
Arguments for a distinct mechanism
Data chunking: exponential rules, difficult to encode/change
Working memory: rule matching grows with number of objects
Slide14Semantic Memory: Big Picture
Supports deliberate storage and retrieval of long-term objects, features, and relations
9 June 2011
Soar's Long-Term Declarative Memory Systems
14
Slide15Architectural Integration
9 June 2011
Soar's Long-Term Declarative Memory Systems
15
Slide16Architectural IntegrationStorage
9 June 2011
Soar's Long-Term Declarative Memory Systems
16
Slide17Architectural IntegrationStorage
9 June 2011
Soar's Long-Term Declarative Memory Systems
17
Slide18Architectural IntegrationStorage
9 June 2011
Soar's Long-Term Declarative Memory Systems
18
Slide19Architectural IntegrationNon-Cue-Based Retrieval
9 June 2011
Soar's Long-Term Declarative Memory Systems
19
Slide20Architectural IntegrationNon-Cue-Based Retrieval
9 June 2011
Soar's Long-Term Declarative Memory Systems
20
Slide21Architectural IntegrationCue-Based Retrieval
9 June 2011
Soar's Long-Term Declarative Memory Systems
21
Slide22Architectural IntegrationCue-Based Retrieval
9 June 2011
Soar's Long-Term Declarative Memory Systems
22
Slide23Mechanism Detail
Encoding
Deliberate (
via rules)Representation: (<long-term identifier> ^attribute value)StorageAgent-modifiable over timeRetrievalAssociative cue: given feature subset, retrieve node (parameterized biases)Expansion: given node, retrieve augmentations
9 June 2011
Soar's Long-Term Declarative Memory Systems
23
Slide24Memory Comparison
Episodic
Semantic
Encoding
AutomaticTop-StateDeliberate
Object+Features/Relations
Storage
No dynamics
Agent modifiable
Retrieval
Disjunctive
Structural
Conjunctive
Features/Relations
9 June 2011
Soar's Long-Term Declarative Memory Systems
24
Slide25Select Prior Work
Efficiency
Towards Efficiently Supporting Large Symbolic Declarative Memories
Nate Derbinsky, John Laird, & Bryan Smith. ICCM (2010)Integrating Learning MechanismsUsing Diverse Cognitive Mechanisms for Action Modeling John Laird, Joseph Xu, & Sam
Wintermute. ICCM (2010)
9 June 2011
Soar's Long-Term Declarative Memory Systems
25
Slide26Today’s Talks
EpMem
SMem
Talkx
Effective and Efficient Historical Memory Retrieval Bias in Soar’s Semantic MemoryNate
Derbinsky
x
Playing with Semantic Memory
Bob
Marinier
x
x
Performance Evaluation of Soar's Declarative Memories
John Laird
x
Do's and
Don't's of Episodic MemoryJustin Lix
x
Supporting Delayed Intentions with Long-Term Memories
Justin
Li
9 June 2011
Soar's Long-Term Declarative Memory Systems
26
Slide27Discussion
9 June 2011
Soar's Long-Term Declarative Memory Systems
27
Slide28Open Research IssuesArchitecture
Functionality
Encoding
StorageRetrievalEfficiencyIntegrationAgent/Task
KnowledgeRepresentationConsistency
SharingCapabilitiesEvaluation
Tasks
Metrics
Tools
9 June 2011
Soar's Long-Term Declarative Memory Systems
28
Slide29Encoding: EpisodicShould
EpMem
capture sub-goal processing?
Space and retrieval time multiplierExplore space saver tradeoff: stop at LTIsSacrifice of fidelity (“changing” episodes)Added deliberation required to “reconstruct” episode contentsCan’t search substructure9 June 2011
Soar's Long-Term Declarative Memory Systems
29
Slide30Example Semantic Knowledge: “soar”
9 June 2011
30
“sense” of the word “soar”
“gloss” with the
synset
-id “200155406”
Soar's Long-Term Declarative Memory Systems
Slide31Example Episode
9 June 2011
31
Soar's Long-Term Declarative Memory Systems
Slide32Pruned Episodic EncodingEpisodic
Semantic
9 June 2011
32
Soar's Long-Term Declarative Memory Systems
Slide33Encoding: SemanticWhere does knowledge come from?
Consolidating from
EpMem
via stability9 June 2011Soar's Long-Term Declarative Memory Systems33
Slide34Structural Stability (1)
9 June 2011
34
Soar's Long-Term Declarative Memory Systems
Slide35Structural Stability (2)
9 June 2011
35
Soar's Long-Term Declarative Memory Systems
Slide36Structural Stability (3)
9 June 2011
36
Soar's Long-Term Declarative Memory Systems
Slide37Structural Stability (4)
9 June 2011
37
Soar's Long-Term Declarative Memory Systems
Slide38Encoding: SemanticWhere does knowledge come from
?
Consolidating from
EpMem via stabilityFrom WM via situational focus (activation)MirroringHypotheticals?9 June 2011Soar's Long-Term Declarative Memory Systems
38
Slide39Storage: EpisodicMemory is cheap, but not unlimited…
Forgetting
Recency
, emotional salienceMergingCan such operations occur in an incremental fashion?Is there a functional benefit to forgetting?9 June 2011Soar's Long-Term Declarative Memory Systems
39
Slide40Storage: SemanticMemory is cheap, but not unlimited…
Forgetting via activation (
ala
ACT-R)Is there a functional benefit to forgetting?9 June 2011Soar's Long-Term Declarative Memory Systems
40
Slide41Retrieval: EpisodicEvaluating retrieval quality
Biases for ambiguous cues
Episode emotional appraisal
Element activationHow to do this efficiently!!??More complex similarity functionsPossibly exploit static structure of IO9 June 2011Soar's Long-Term Declarative Memory Systems
41
Slide42Retrieval: SemanticEffective & efficient biases (
ala
ACT-R)
Base-levelSpreading (more than 1-level?)Partial matchNoiseOther?9 June 2011Soar's Long-Term Declarative Memory Systems
42
Slide43EfficiencyAsynchronous retrievals
Pragmatic: max time per cycle
Explore multiple approaches to a problem, while LTMs are searching
Faster memory -> better task performance?9 June 2011Soar's Long-Term Declarative Memory Systems
43
Slide44IntegrationLearning to control LTMs
Knowledge-level learning/retrieval strategies
Symbolic vs. Perceptual LTMs
How can symbolic LTM inform perceptual processes?9 June 2011Soar's Long-Term Declarative Memory Systems
44
Slide45KnowledgeWhat types of representations do we need to support and learn large, diverse semantic stores?
How do we maintain consistency between the long-term memories (or at least remain robust to inconsistencies)?
To what extent can long-term knowledge be shared between agents?
How do we capture high-level capabilities as re-usable components (i.e. rule libraries)?9 June 2011Soar's Long-Term Declarative Memory Systems
45
Slide46EvaluationWhat is a “comprehensive” set of domains/tasks by which to measure progress of LTM development?
M
etrics?
At what time scale should we evaluate LTMs for generally intelligent agents?Is this feasible in real-time domains?9 June 2011Soar's Long-Term Declarative Memory Systems
46
Slide47ToolsAs agents persist for long periods of time and build up their own knowledge stores, how do we assess/visualize progress and debug agents?
9 June 2011
Soar's Long-Term Declarative Memory Systems
47