A Soar s Eye View of ACT-R

A soar s eye view of act r
1 / 17
1059 days ago, 425 views
PowerPoint PPT Presentation

Presentation Transcript

Slide 1

A Soar's Eye View of ACT-R John Laird 24 th Soar Workshop June 11, 2004

Slide 2

Soar/ACT-R Comparison What changes in respect to ACT-R would essentially modify Soar? Extensions (enactment, RL, EpMem) as well as principal changes. What changes in respect to Soar would essentially adjust ACT-R? Take off Soar ACT-R ACT-R ACT-R

Slide 3

Obvious Similarities Soar 9 ACT-R 5 Input/Output Buffers & Async. Buffers & Async. Fleeting memories Graph Structure Chunks in supports Activation Base Activation Long-term memories Production Rules Production Rules Episodes Declarative Memory Rule Utilities Chunk Associations Sequential control Operator Production Goal Structures State stack Goal & Declarative Memory Learning Chunking Production Composition Reinforcement Utility Learning Episodes Chunks - > Decl. Memory Goal & Chunk Association Base Activation

Slide 4

Soar Unbounded chart structure Multi-esteemed characteristics: sets Decision on ^operator of state I-support and o-bolster Explicitly speak to state Short-term identifiers Generated every time recovered Values can be long haul images ACT-R Chunks (level structures) in cushions One lump/cradle Chunk sorts with altered spaces Goal, Declarative Memory, Perception All constant until supplanted/adjusted Long-term identifiers for every lump Provides various leveled structure state perception objective revelatory memory #3 recognition #45 red #3 "x" #9 Short-term Memories

Slide 5

Implications for Soar Unbounded working memory No simple approach to move subset of transient memory to long haul memory piece by piece Can't keep up associations between articles without long haul memory images Makes it conceivable to decide comes about naturally Supports programmed evacuation of immaterial information state

Slide 6

Implications for ACT-R Bounded representation Long-term memory images permit dynamic epitome Can figure out how to test just piece id rather than substructure Flat representation Hard to speak to sets Requires "unloading" of protest images to get to highlights But can learn decides that get to images specifically How would it be able to perceive organized items from discernment? (Mixing?) Unitary question representation supremacy (versus free components) All elements are similarly essential (enactment is question based) Chunk sorts are structurally important revelatory memory objective #3 observation #45

Slide 7

Implications for ACT-R II Persistence Easy to have conflicting convictions Consistency dependably contends with other thinking Working Memory = recovered LTM Declarative Memory (Changes in working memory change explanatory memory) No memory of old values in lumps Difficult to keep up autonomous duplicates of same protest Hypothetical thinking definitive memory objective #3 recognition #45

Slide 8

Fundamental Issue: Long-Term Object Identity Architectural (ACT-R) versus Information based (Soar) Connecting to discernment Connecting to other long haul recollections Copying structures

Slide 9

Decision Making Soar ACT-R Generate features Parallel rules Sequential rules Generate alternatives Parallel rules Match manage conditions Compare & rate alternatives Parallel rules Rule utility Select Architecture Architecture Apply Parallel rules Rule actions Dimensions for correlation: Simple measurements # of thinking steps required # of consecutive administer firings # information units (rules) required ACT-R regularly exchanges off lumps + understanding + learning for principles. Capacities Expressibility Use setting Open to meta-thinking Modification through learning

Slide 10

Execution Steps Soar ACT-R Generate highlights (F) Parallel rules Sequential rules Generate alternatives (O) Parallel rules Match administer conditions Compare & rate choices (C) Parallel rules Rule utility Select Architecture Architecture Apply (A) Parallel rules Rule activities # of manage firings F + O + C + A F + 1 # of consecutive steps 1 F + 1 This is convoluted by definitive memory recoveries in ACT-R – however they are not by any stretch of the imagination procedural information specifically included in basic leadership, in spite of the fact that they are some of the time included in a roundabout way.

Slide 11

Propose and Apply Knowledge Units For a solitary O that can be chosen in S Situations and has A was of Applying: Soar: O + A tenets ACT-R: O * A guidelines O: Independent Proposals An: Independent Applications Op

Slide 12

Q i Q k Q j Selection Knowledge Units In Soar, autonomous numeric impassive principles consolidate values for choice Allows direct mixes of attractive quality In ACT-R, just a solitary utility esteem is connected with every administer No run time mix Conflates lawfulness (proposition) and allure Must have isolate control for every novel setting application match Architecture Architecture

Slide 13

Expressibility Soar permits "open choices" Which learning contributes is decide at run time Does not require pre-aggregation of vital elements. Isolates information about "can" do an activity from "ought to" Makes simple to express and add information to adjust strategy Symbolic inclinations Possibility of one-shot learning for basic leadership Can be advised not to do an activity (and defeat measurable) Can figure out how to not do an activity

Slide 14

Use Run-time Context Soar ACT-R Generate alternatives Yes – rules Yes – administer conditions Compare/rate alternatives Yes – rules No – govern utility Select Architecture Architecture Apply Yes – rules No – manage activity

Slide 15

Meta-Reasoning Soar has tie impasses & subgoals Can identify when learning is unverifiable/fragmented Can utilize discretionary thinking to examine and settle on choice Including look-ahead arranging with theoretical states Can return comes about that adjust the choice Learning can straightforwardly alter choice ACT-R Difficult to recognize vulnerability a & reason about choice Could make impasse when utilities are close or dubious Difficult to change choice without experience How could other thinking change a creation run choice?

Slide 16

Predictions! ACT-R Something to manage meta-discernment Detecting vulnerability and ponder thinking to manage it (and the learning). Arranging Integration of feeling/agony/joy for learning Episodic memory Soar Long-term revelatory memory & compositional definitive adapting Some one will construct ASCOT-ARR! ACT-R memory structure with Soar administrators

Slide 17

Gold and Coal Goal: Having elective designs Provides motivation for compositional adjustment Provides correlation Forces us to look at self-assertive choices Coal: Most correlations with date are: Informal, (for example, this) Not hypothesis coordinated (AMBER) Confound programming & engineering Not precisely same assignment