discussant for last session bob balzer teknowledge architectural issues
DESCRIPTION
Architectural Comparison 3T World 1/10 sec. limited state limited projection memory interpreter task1 subtask task2 task3 1 sec. memory of immediate actions no projection 10s sec. persistent state and choices projection Adaptive Mission Planner Controller Synthesis Module Real Time System CIRCA Deliberation scheduling Planning ExecutionTRANSCRIPT
Discussant for Last Session
Bob BalzerTeknowledge
Architectural Issues
Pat LangleySome motivating assumptions
• We should move beyond isolated phenomena and capabilities to develop complete intelligent systems
• AI and cognitive psychology are close allies with distinct but related goals• Characterize intelligent behavior at the level of functional structures and
processes, not at the implementation or knowledge levels• Cognitive architecture should make commitments to representations and
organizations of knowledge, the memories in which such knowledge resides and the processes – performance and learning – that operate upon them
• A cognitive architecture specifies the infrastructure that does not change over domains and time, as opposed to knowledge, which does vary
• A cognitive architecture should have an associated programming language for encoding knowledge and constructing intelligent systems
• An architecture should demonstrate generality and flexibility rather than success on a single domain or application
components, or processes, which do vary APIs
, adding components, and modifying processes
Architectural Comparison3T
World
1/10 sec. limited state limited projection
memory interpreter
task1 subtask subtasktask2task3
1 sec. memory ofimmediate actions no projection
10s sec. persistent stateand choices projection
Adaptive Mission Planner
Controller Synthesis Module
Real Time
System
CIRCA
Deliberation scheduling
Planning
Execution
Architecture CompositionHomogeneous Composition
Roles, Goals
Real-Time Reactions
Planned Actions,Planned Negotiations
Adaptive Mission Planner
Controller Synthesis Module
Real Time
System
Adaptive Mission Planner
Controller Synthesis Module
Real Time
System
Extending Performance Guarantees to Multi-Agent Teams
Heterogeneous Composition
Architecture CompositionHomogeneous Composition
path planner
simulationIVHM
scheduler
Planner
Planner
Planner
T3 Multi-Agent Teams
Architecture Composition Heterogeneous Composition
Soar
Interact with a complex world - limited uncertain sensingRespond quickly to changes in the world Use extensive knowledgeUse methods appropriate for tasksGoal-drivenMeta-level reasoning and planningGenerate human-like behaviorCoordinate behavior and communicate with others Learn from experienceIntegrate above capabilities across tasksBehavior generated with low computational expense
Target Application Behavioral Capabilities
FewProblemsMatchProfile}
CognitiveArchitecturei
CognitiveArchitecturei
CognitiveArchitecturei
Special Purpose Cognitive Architectures
ArchitectureSelector
Architecture Composition Heterogeneous Composition
(from Jonathan Grach)
• Soar as a component in a larger architecture
Speech Recognition (HTK)
Semantic Parser
Motion/ Gesture Scheduler (Beat)
Text to Speech (Festival)
World Simulator
Animation System
BDI
Haptek
Com
mun
icat
ion
Bus
Audio (Protools)
Voice Input
Vega
Projection System
Speakers (10.2)
Soar Planning
DialogueAction Selection
Perc
eptio
n
NLG
Emotion
NLU pragmatics
Child Healthy:False
AccidentIntend: FalseBlame: unresolved
Assist Eagle 1-6:False
Eagle 1-6 AssistDesire: LT
Belief: False
Child-HealthyDesire: SGTBelief: False
Probability: 75%
Get MedevacResponsibility:LTIntend: True
Medevac Available:True
Past FuturePresent
Cognitive Representation
Soar’s Working Memory
Planning Perception Dialogue Action
Soar operators
• Defining architecture (components) within Soar– Accomplished through escape mechanism (no extension
API)
Comparative Framework• Representational elements
– Inputs, Justified Beliefs, Assumptions, Desires, Active Goals, Plans, Actions, Outputs
• Design dimensions– Representation formalism
• How is each type of element represented?– Commitment strategy
• Under what conditions does each type of element get selected/activated/instantiated?
– Reconsideration strategy• Under what conditions does each type of element get
removed/deactivated/released?
Thinking inside the box
Missing Elements
• Deliberate attention• Parallel active goals• Resources and limitations• Multi-agent/social elements• Learning• Episodic memory
Thinking inside the box
Intermediate, Reusable Components
• Vocabulary for defining components• Uses: Modeling, Comparative, Generative
T h i n k i n g i n s i d e t h e b o xM a r c h 2 2 - 2 3 , 2 0 0 3 A r c h i t e c t u r e W o r k s h o p S l i d e 1 7
W o r k i n g M e m o r yE l e m e n t
B e l i e f
A s s u m p t i o n
D e s i r e
G o a l
E n t a i l m e n t
A c t i v a t e da s s e r t i o n
T h i n k i n g i n s i d e t h e b o xM a r c h 2 2 - 2 3 , 2 0 0 3 A r c h i t e c t u r e W o r k s h o p S l i d e 1 8
E n t a i l m e n t
C r e a t e : T r u t h m a i n t e n a n c e
R e l e a s e : T r u t h m a i n t e n a n c e
A s s u m p t i o n
C r e a t e : P e r s i s t e n t a s s e r t i o n
R e l e a s e : P e r s i s t e n t a s s e r t i o n
A c t i v a t e dA s s e r t i o n
C r e a t e : A c t i v a t i o n t h r e s h o l d
R e l e a s e : A c t i v a t i o n t h r e s h o l d
T e m p o r a lA s s u m p t i o n
C r e a t e : P e r s i s t e n t a s s e r t i o n
R e l e a s e : D e c a y f u n c t i o n
C o n t e x t - S e n s i t i v eA s s u m p t i o n
C r e a t e : P e r s i s t e n t a s s e r t i o n
R e l e a s e : P e r s i s t e n t a s s e r t i o nR e l e a s e : T r u t h M a i n t e n a n c e
One method for extending architectures: Libraries
Architectural Schema
• Generative grammar for a class of architectures– Different architectures include mechanisms in different subsets of the boxes– Different possible information links, – Different possible control relationships.– Also differences in forms of representation and types of mechanisms.
Soar ++
• How new capability was added to Soar– By modifying system (and done by system developers)
• How could such capabilities be added “externally”– What architectural extension mechanism are provided
Episodic Learning[Andrew Nuxoll]
• What is it?• Not facts or procedures but memories of specific events• Recording and recalling of experiences with the world
• Characteristics of Episodic Memory• Autobiographical• Not confused with original experience• Runs forward in time• Temporally annotated
• Why add to Soar architecture? • Not appropriate as reflective learning • Provides personal history and identity• Memories that can aid future decision making & learning• Can generalize and analyze when time and more knowledge are available
Reinforcement Learning[Shelley Nason]
• Why add it to Soar?• Might capture statistical regularities automatically/architecturally• Chunking can do this only via deliberate learning
• Why Soar?• Potential to integrate RL with complex problem solver• Quantifiers, hierarchy, …
• How can RL fit into Soar?• Learn rules that create numeric probabilistic preferences for operators• Used only when symbolic preferences are inconclusive• Decision based on all preferences that are recalled for an operator
• Why is this going to be cool?• Dynamically compute Q-values based on all rules that match state• Get transfer at different levels of generality
Architectural Extension
Distinctive Features of EPIC Work
• Emphasis on executive processes that coordinate multitask performance– Multitask performance stresses the architecture.– An important but underdeveloped area for theory.
• Take advantage of underexploited but powerful constraints:– Perceptual-motor abilities and limitations.– Detailed and exact quantitative fits to human data.
• “Zero-based” theoretical budget:– Question traditional assumptions.– Do not add a mechanism until it is needed to account for data.– Avoid egregious assumptions of cognitive limitations.– Prefer strategy limitations over architectural ones.
• Focus on major phenomena and mechanisms that are important determinants of performance, rather than minor “interesting” ones.
• Compare multiple strategies for doing a task.– Isolate strategy effects from architectural properties.
Shouldn’t stress performance architecture
Builds and repairs fully-detailed flight schedules for any planning horizon, without losing sight of command objectives,
providing new opportunities to explore and manage alternative futures, in 1/10th-1/100th of current time
o ConstraintsConstraintso Training code pre-requisitesTraining code pre-requisites
from T&R Manualfrom T&R Manualo Fly dayFly dayo Day & night missionsDay & night missionso Crew day rulesCrew day ruleso Turn-around & briefing timeTurn-around & briefing timeo Instructor requirementsInstructor requirementso Range capabilitiesRange capabilities
o Availability & suitabilityAvailability & suitabilityo Merging and splittingMerging and splitting
o Range boardRange boardo Pilot SNIVELsPilot SNIVELso Aircraft availabilityAircraft availabilityo Simulator schedule Simulator schedule
Range UseRange Use
Pilots’ ViewPilots’ View
Scheduling OfficerScheduling OfficerFeedbackFeedback
Status of SNAP: Schedules Negotiated by Agent-Based Planners
Identifies needed ranges
Tracks pilots
Compares results to guidance
Knows the situation
Accepts guidance at any level of specificity
Lets users adjust priorities
InputsInputs
OutputsOutputs
Prioritized GuidancePrioritized Guidanceo Squadron focusSquadron focus o Pilot focuso Pilot focus o Sortie cycleo Sortie cycleo Pilot buildsPilot builds o Pilot specific training codeo Pilot specific training code o Fly dayo Fly dayo Pilot snivelsPilot snivels o Rangeso Ranges o No. aircraft of each typeo No. aircraft of each type
Obeys the law
SNAP Agents: Trade-off Exploration,SNAP Agents: Trade-off Exploration, Win-Win Scheduling Solutions Win-Win Scheduling Solutions
Flow managerFlow manager PilotsPilots AircraftAircraft MissionsMissions RangesRanges PMCFPMCF SimulatorsSimulators Sim. MonitorsSim. Monitors ODOODO OrdnanceOrdnance AcademicsAcademics
Yearly Training PlanYearly Training PlanTEEPTEEP
Flight Hour ProgramFlight Hour Program
Monthly Training PlanMonthly Training Plan
Weekly Training PlanWeekly Training Plan
Daily Flight ScheduleDaily Flight Schedule
Active Flight ScheduleActive Flight Schedule
Daily ScheduleDaily Schedule
Produces schedules
Weekly Sched.Weekly Sched.
Monthly Sched.Monthly Sched.
ElectronicElectronicFeed toFeed to
MaintenanceMaintenance
Architectural Challenges• Scaling
– Three Guarantees1. No Single Cognitive Architecture will handle all problems
Heterogeneous collection of Cognitive Architectures2. Not all activity will occur within Cognitive Architecture
Cognitive Architecture must interface with COTS & GOTS systems3. Your Cognitive Arch. will be a component in some bigger system
Cognitive Architecture must operate in an imposed architecture• Architectural Extension
– Identify points of variabity– Provide enumerated choices across that variability and procedural alternatives
• Heterogeneous Composition– API for embedding one architecture as a component within another– Sharing/adding Knowledge across architectures
or across heterogeneous components• Handling uncertainty• Security