A Different Kind of KR COMP34512 Sebastian Brandt brandt@cs.manchester.ac.uk Friday, 2 May 2014
Varieties of KR • We’ve been looking at a fairly narrow set – Mostly (first order) logic based – “Ontology” languages • Primarily “conceptual” knowledge – Good for terminologies other data management • Wide range of possibilities – Other than first order logics • Modal, probabilistic, temporal, spatial, nonmonotonic, etc. – Code! – Semantic nets (SKOS), production rules, Bayesian networks – Etc. etc. etc. • Often determined – By needs of the domain 2 Friday, 2 May 2014
Representational Domain • Recall cognitive adequacy – Weak ≈ usability – Strong ≈ how we think • A case where SCA is (perhaps) required – When we want to represent human minds! • Why? – To test (or enact) psychological theories – Cognitive Tutors – Usability testing – Human-like intelligences?!??! 3 Friday, 2 May 2014
Cognitive Architectures • Idea – A general, typically comprehensive, model of cognition • Or the foundations of cognition • Or basic structure of cognition • A unifying model of cognition • Tools for building cognitive/intelligent agents – At least, psychologically plausible • Ideally, psychologically accurate! • Many proposed and used – Symbolic • ACT-R, SOAR, EPIC, etc. • Thought ≈ manipulation of symbols • “Language of thought” – Connectionist • E.g., Neural networks, computational neuroscience (SpiNNaker) • Thought ≈ patterns of activation in simple connected units 4 Friday, 2 May 2014
CogArchs as KR formalisms • Symbolic Cognitive Architectures – Often have a “language” • Symbols and symbol manipulation after all – Syntax • Usually clear – Semantics • Not so clear – Reasoning procedure • Also not always clear • They face similar design tradeoffs – Expressivity, usability, computability 5 Friday, 2 May 2014
ACT-R • “Adaptive Control of Thought—Rational” – Several specialized modules • Visual module, long term memory, manual module, goal module, etc. – Central production system coordinates • “Production” rules push and pull information between modules – Mapping onto neural substrates • Other biological constraints • E.g., 50 ms “cognition cycle”’ – Historical bias • toward high level cognition – Address many aspects • Acquisition, learning • “Compilation” • Decay • Error http://act-r.psy.cmu.edu/papers/526/FSQUERY.pdf 6 Friday, 2 May 2014
Schematic Overview http://act-r.psy.cmu.edu/about/ 7 Friday, 2 May 2014
ACT-R Kinds of Knowledge • Declarative – Articulable knowledge; propositional – Represented as “chunks” • I.e., typed sets of key-value pairs (i.e., “slots”) • A weak sort of object – Stored in the declarative memory • “Viewed” through the “retrieval buffer” • Limited access! Must “recall” knowledge! • Procedural – Actionable “ways of doing things” – Production rules – Stored in procedural memory – NOT directly retrievable • Triggered by goals and retrieved knowledge (chunks “in mind”) • Manipulates goals, retrieves knowledge, alters knowledge – Both declarative and procedural! 8 Friday, 2 May 2014
Counting example • How to capture counting from 2 to 4 • We need to have declarative knowledge – That after 2, comes 3, and after 3, comes 4 • We need to have procedural knowledge – That to “count from X to Y”, we • Start at X • Go to the next number, N • If N = Y, we made it! • Otherwise, repeat with X = N 9 Friday, 2 May 2014
Counting example: Declarative (chunk-type count-order first second) (chunk-type count-from start end count) (add-dm (b ISA count-order first 1 second 2) (c ISA count-order first 2 second 3) (d ISA count-order first 3 second 4) (e ISA count-order first 4 second 5) (f ISA count-order first 5 second 6) (first-goal ISA count-from start 2 end 4)) (b ISA count-order first 1 second 2) Is this psychologically realistic? From http://act-r.psy.cmu.edu/tutorials/ 10 Friday, 2 May 2014
Counting example: Procedural (P increment If =goal> The goal is to ISA count-from count from count =num1 our current count - end =num1 but we’re not done =retrieval> and we have “in mind” ISA count-order in the normal order first =num1 given where we are second =num2 what comes next ==> Then =goal> Change the goal count =num2 to count from the next nr +retrieval> and recall ISA count-order What comes next first =num2 after our current count !output! (=num1) (Count out loud!) ) Is this what we do when counting? From http://act-r.psy.cmu.edu/tutorials/ 11 Friday, 2 May 2014
Counting example: Setup ... • We focused on the counting (first-goal – Not on what would cause us to count! ISA count-from start 2 • “Would you please count aloud from 2 to 4!” end 4)) – But we still need to initiate action • So we gin up some bits (goal-focus first-goal) (p start =goal> ISA count-from start =num1 (P stop count nil =goal> ==> ISA count-from =goal> count =num count =num1 end =num +retrieval> ==> ISA count-order -goal> first =num1 !output! (=num) ) ) 12 From http://act-r.psy.cmu.edu/tutorials/ Friday, 2 May 2014
Counting! (run 1) 0.000 GOAL SET-BUFFER-CHUNK GOAL FIRST-GOAL REQUESTED NIL 0.000 PROCEDURAL CONFLICT-RESOLUTION 0.000 PROCEDURAL PRODUCTION-SELECTED START 0.000 PROCEDURAL BUFFER-READ-ACTION GOAL 0.050 PROCEDURAL PRODUCTION-FIRED START 0.050 PROCEDURAL MOD-BUFFER-CHUNK GOAL 0.050 PROCEDURAL MODULE-REQUEST RETRIEVAL 0.050 PROCEDURAL CLEAR-BUFFER RETRIEVAL 0.050 DECLARATIVE START-RETRIEVAL 0.050 PROCEDURAL CONFLICT-RESOLUTION 0.100 DECLARATIVE RETRIEVED-CHUNK C 0.100 DECLARATIVE SET-BUFFER-CHUNK RETRIEVAL C Ok, a demo would 0.100 PROCEDURAL CONFLICT-RESOLUTION 0.100 PROCEDURAL PRODUCTION-SELECTED INCREMENT 0.100 PROCEDURAL BUFFER-READ-ACTION GOAL be easier! 0.100 PROCEDURAL BUFFER-READ-ACTION RETRIEVAL 0.150 PROCEDURAL PRODUCTION-FIRED INCREMENT 2 0.150 PROCEDURAL MOD-BUFFER-CHUNK GOAL 0.150 PROCEDURAL MODULE-REQUEST RETRIEVAL 0.150 PROCEDURAL CLEAR-BUFFER RETRIEVAL 0.150 DECLARATIVE START-RETRIEVAL 0.150 PROCEDURAL CONFLICT-RESOLUTION 0.200 DECLARATIVE RETRIEVED-CHUNK D 0.200 DECLARATIVE SET-BUFFER-CHUNK RETRIEVAL D 0.200 PROCEDURAL CONFLICT-RESOLUTION 0.200 PROCEDURAL PRODUCTION-SELECTED INCREMENT 0.200 PROCEDURAL BUFFER-READ-ACTION GOAL 0.200 PROCEDURAL BUFFER-READ-ACTION RETRIEVAL 0.250 PROCEDURAL PRODUCTION-FIRED INCREMENT 3 0.250 PROCEDURAL MOD-BUFFER-CHUNK GOAL 0.250 PROCEDURAL MODULE-REQUEST RETRIEVAL 0.250 PROCEDURAL CLEAR-BUFFER RETRIEVAL 0.250 DECLARATIVE START-RETRIEVAL 0.250 PROCEDURAL CONFLICT-RESOLUTION 13 0.250 PROCEDURAL PRODUCTION-SELECTED STOP Friday, 2 May 2014
A Taxonomic Example • Let’s consider a familiar (in this class!) task – Reasoning about categories • Their hierarchical relations • Their relational structure moves animal skin wings gills bird fish flies swims edible yellow can't fly dangerous salmon ostrich shark canary tall swims swims sings From http://act-r.psy.cmu.edu/tutorials/ 14 Friday, 2 May 2014
Taxonomy: Declarative moves (chunk-type property animal skin object attribute value) wings gills bird fish (chunk-type is-member flies swims object category edible yellow can't fly dangerous salmon ostrich judgment) shark canary tall swims swims sings (p1 ISA property object shark attribute dangerous value true) (p1 ISA property (p2 ISA property object shark attribute locomotion value swimming) object shark (p3 ISA property object shark attribute category value fish) attribute dangerous ... value true) (p7 ISA property object fish attribute breathe value gills) (p8 ISA property object fish attribute locomotion value swimming) (p9 ISA property object fish attribute category value animal) (g1 ISA is-member object canary (g1 ISA is-member object canary category bird judgment nil) category bird (g2 ISA is-member object canary category animal judgment nil) judgment nil) (g3 ISA is-member object canary category fish judgment nil) From http://act-r.psy.cmu.edu/tutorials/ 15 Friday, 2 May 2014
Taxonomy: Setup From http://act-r.psy.cmu.edu/tutorials/ 16 Friday, 2 May 2014
Taxonomy: Procedural -- Direct From http://act-r.psy.cmu.edu/tutorials/ 17 Friday, 2 May 2014
Taxonomy: Procedural -- Indirect From http://act-r.psy.cmu.edu/tutorials/ 18 Friday, 2 May 2014
Recommend
More recommend
