ACT-R & A 1000 Flowers ACT-R Adaptive Control of Thought Rational Theory of cognition today Cognitive architecture Programming Environment 2 Core Commitments of the Theory Modularity (and what the modules are) Distinction of declarative and procedural knowledge Two types of processing: symbolic & sub-symbolic Constrained parallelism Core Components of the Architecture Chunks Modules Buffers Productions Parameters & timing constants 3 4 Chunks Represent declarative knowledge Take the form of slot-value lists Example: London is north of Paris chunk-name!!ob1!!london!!ob2!!paris!!relation!north-of! Chunk values can be chunks At each point in time, each chunk has a certain activation 5 Modules Parts of the architecture that provide certain (cognitive) capabilities Modules are largely independent of each other (parallelism) Modules are encapsulated (Nearly) Each module has one or more buffers 6 1
Buffers Each buffer is associated to exactly one module A module s buffer(s) provides the only interface to that module Each buffer can contain no or exactly one chunk (serial bottleneck) => To interface with modules all information has to be translated into chunk form Productions If-then rules (as in production systems) LHS and RHS refer to buffers LHS checks buffer contents RHS modifies buffer contents At each point in time, each production has a certain utility 7 8 Time Constants & Parameters Modules in Vanilla ACT-R Operations within modules take (simulated) time Default times specified for operations A large part of ACT-R s functionality can be controlled by parameters Use of modules Use of (parts of) processing types Processing styles Processing times Etc. 9 Goal Imaginal Procedural Motor Vision BOLD Audio Speech Temporal 10 Module Structure ACT-R as Production System 11 Select execute cycle as in any PS Match LHS of productions to buffer content => conflict set of applicable productions Conflict resolution: highest utility Buffers constitute data base No direct exchange between modules Information exchange only through procedural module 12 2
Modules in Vanilla ACT-R BOLD (simulate fmri) Goal Audio (auditory attentional system) Imaginal Speech (speaking capabilities) Procedural Temporal (estimating short time spans) Motor Vision Motor Module Provides ACT-R with hands Manual buffer Main supported action: Pressing keys on a standard keyboard Also: Moving and clicking a mouse Includes calculation of times required for requested actions 13 14 Vision Module Realizes visual attention not perception Two buffers: visual-location, visual Visual buffer holds obect information (what) Visual-location holds location information (where) Normal operation: Request location Vision Module Shift attention to location (request obect at location) Use obect information Provides times required for attention shifts and obect encoding 15 16 Goal / Imaginal Task: Add 64 and 33. Subtasks required Add digits in ones place Store result, identify carry (if any) Add digits in tens place (+ carry) Store result, identify carry (if any) Report answer Goal / Imaginal Task: Add 64 and 33. To solve the task it is necessary To keep track of the current / next subtask (control state) Intermediate results (problem state) Previously, control and problem state were both managed through the goal module / buffer. ACT-R 6.0 has different modules for these 17 18 3
Goal / Imaginal Imaginal: Stores problem state imagination delay Goal: Stores control state Realizes control flow As coordinating productions through DB Both modules basically not specified 19 knowledge base of ACT-R Knowledge stored as chunks Chunks available through retrieval requests To retrieve a chunk its activation has to be above the retrieval threshold τ Activation of chunk i,, is given as +! W S i 20 +! W S i +! W S i B i is the base-level activation Roughly: Past usefulness of chunk n!d B i = ln( " t ) where t = time since th practice =1 More recent formulation: ε is a noise term Noise is computed by sampling from a logistic distribution Adds short term inhibition 21 Standard deviation is set through parameter s Noise allows for non-deterministic behavior! 2 = " 2 3 s2 22 +! W S i Sum represents contextual influence Retrieving a chunk takes time Time depends on chunk activation Time i = Fe! When no chunk is above threshold Retrieval failure Context usually restricted to chunk in goal / imaginal buffer 23 Time is computed with = τ Probability of retrieving a chunk 24 4
Procedural Precedence of matching productions is based on utility Utility U i of production i, is governed by U i (n) =U i (n!1)[r i (n)!u i (n!1)] where α is the learning rate During selection noise is added to the current utility of each production Utility noise is as activation noise Procedural U i (n) =U i (n!1)[r i (n)!u i (n!1)] R i is the reward for production i Rewards are triggered by productions or other events (set manually) The reward actually received by production i depends on the time lag between reward and production firing 25 26 Procedural Probability of a matching production being executed Pr obability(i) =! Utility does not influence execution time of selected production eu i / 2s e U / 2s Modularity & Flowers Modularity eases adding new capabilities to ACT-R In fact, constant extension / modification is seen as desirable by the people behind ACT-R Let a thousand flowers bloom! i.e., let existing ideas contend such that the best may prevail 27 28 Toothbrushes & Theories What do toothbrushes and theories have in common? 1) Everyone wants an own 2) No one wants to use the ones of other people ACT-R Toothbrushes It is hard to identify something such as the ACT-R Tendency of people to extend / modify ACT-R to their own needs Plus: Large DoF through parameterization of ACT-R Everyone is bringing his / her own toothbrush into ACT-R 29 30 5
Some facts ACT-R allows to implement arbitrary Turing complete computing formalisms Virtually no two ACT-R models are using the same theory. => Is ACT-R a cognitive architecture in Newell s sense? 31 6