ACTRPM

1
ACT-R/PM

• CS/ISYE/PSYC 7790
• Fall 2003

2
Administrivia

• This week ACT-R/PM
• Byrne Anderson
• Byrne
• Lab ACT-R/PM
• Next week Preattentive factors
• Treisman Gelade
• Schneider Shiffrin

3
What weve looked at so far

• ACT-R
• Preattentive factors and visual search
• Models of visual search
• Guided search (Wolfe, others)
• SERR
• Today ACT-R/PM

4
Outline

• ACT-Rs model of vision
• ACT-R/PMs EPIC saga
• Other models of vision within this framework

5
ACT-Rs visual interface

• Limited to interactions involving a computer
  screen
• Wherever attention is shifted, declarative chunks
  can be generated

6
Example of an ACT-R encoding

• Visual chunks encoded as chunks of type
  VISUAL-OBJECT with Cartesian coordinates
• Object
• isa VISUAL-OBJECT
• screen-position (125 100)
• value H

7
Feature-based model of attention

• Three types of information can be used to attend
• Particular locations and directions
• Particular feature types (e.g., color)
• Features that have not yet been attended
• Not a full model of attention, but allows
  modeling some attention characteristics
• Any feature type can be used (seemingly), and
  feature conjuncts (pink and vertical) can be
  attended to, even though this is harder for
  humans (Triesman Gelade, 1980)
• Can limit processing to unattended features, but
  no intrinsic model of inhibition of return shown
  by humans

8
Examples of the visual interface

• Sperling task
• Subitization task
• Menu selection

9
Sperling task (1960)

• 3x4 grid of letters
• Whole report
• 50 msec masked presentation
• 4.4 letters reported, on average
• Partial report
• Audio cue for row
• Could report back 3.3 letters per row
• Delayed report
• 2.5 letters at .12 s
• 2.0 letters at .4 s
• 1.5 letters at 1.0 s

10
ACT-R model of Sperling task

• Encode-screen and Encode-Row production rules
• Assume an iconic memory that lasts approximately
  4.4 times the length of a production firing
• Thus 4.4 items for whole report condition
• Approximately 810 msec.
• For partial report, assume that some digits may
  already be encoded, and that remaining attention
  is directed to unattended digits on the given row
• Use 810 msec figure to limit processing
• Fit to data is good

11
Results for Sperling Task
Anderson Lebiere, 1998
12
ACT-R Model of Subitizing

• Covered previously
• Subitizing productions
• See-one
• See-two
• See-three
• Counting productions
• Stop and start productions
• Fit to data to get two-stage result

13
Word superiority effect

• Word superiority effect
• Letters are easier to recognize in the context of
  words
• WORD ? (D or K?)
• Examined by McClelland and Rumelhart
• Set of letter features
• Set of words
• Used set of prior and conditional probabilities
  to choose best word or letter
• Anderson created ACT-R model using same set of
  letter features
• Demonstrated that he could model word superiority
  by changing size of attentional set
• WOR_ ? WORD

14
Applying Word superiority to menu selection

• Nilsen model of menu selection
• Selecting an item in a menu
• 103 msecs/item, starting from topmost item in
  menu
• Linear function
• But, what about the confusability of the menu
  items?
• Modeled using McClelland and Rumelhart data
• Select either letter or digit on background of
  letters or digits
• ACT-R predicts, and experiments showed, that
  different selection/background combinations
  distracted at different levels
• Number-on-number condition had the greatest
  feature overlap

15
Outline

• ACT-Rs model of vision
• ACT-R/PMs EPIC saga
• Other models of vision within this framework

16
ACT-R/PMs EPIC saga

• Model Human Processor (Card, Moran Newell,
  1983)
• Resources arranged by modality
• Vision, hands, etc.

17
One alternative approach to MHP modeling GOMS

• Parts
• Goals
• Operators
• Methods
• Selection rules
• Arranged by modality
• Output as PERT chart
• Milliseconds instead of days
• Modalities instead of workers
• Good for displaying to middle managers

18
How GOMS modeling is typically done

• Set of templates for common tasks in planning
  software
• Templates are filled in, arranged, and linked by
  hand
• Apex/GOMS, which Ive mentioned before, automates
  this process
• Templates have default values
• Automatic arrangement of set of tasks
• Can model multitasking
• Click on a word while visually searching for a
  menu item
• Planning is done using a reactive planner

19
EPIC
20
EPIC (Kieras Meyer, 1996)

• Model of multimodal processing
• Auditory and visual processor for input
• Ocular, vocal, and motor output
• Resolution of 50 msec.
• Production rule system
• Data memory
• Production rule memory

21
Limitations of EPIC

• No model for iconic memory
• No limitations on which productions can fire
• Simultaneous addition and multiplication
  functions possible

22
ACT-R/PM

• Combines EPIC and the ACT-R visual interface
• From ACT-R visual interface
• Richer production system
• Better memory model
• From EPIC
• Better motor control model
• Parallel system operation
• Speech and audition timing parameters

23
ACT-R/PM Architecture

• Picture from p. 173
• Cognition layer
• Production memory
• Declarative memory
• Perceptual/Motor layer
• Visual module
• Motor module
• Speech module
• Audition module

24
Utilizing the modules

• Each module has two or more buffers that are
  available to ACT-R
• Accessed using operator
• Each has a module state buffer that indicates
  whether the module is free or busy
• Accessing a module when it is already busy is
  known as jamming, and it generates an error
  message
• Vision module buffers
• Visual, visual-location, visual-state

25
Example from visual buffer Attend-letter
production rule
Has a visual location been attended?

• (P attend-letter
• goalgt
• ISA read-letters
• state find-location
• visual-locationgt
• ISA visual-location
• visual-stategt
• ISA module-state
• modality free
• gt
• visualgt
• ISA visual-object
• screen-pos visual-location
• goalgt
• state attend
• )

Is the visual module currently available?
Then store attended object in visual buffer
26
Visual module

• Models of attention
• Spotlight (Posner, 1980)
• Feature synthesis (1990)
• Guided search (Wolfe, 1994)
• Visual scene parsed into high-level visual
  features
• Attention can be directed within iconic memory
• Can discriminate between current and past state
  of the visual system
• (Evidently by comparing the current percept with
  that in iconic memory?)
• Can handle movement and change

27
Motor Module

• Runs in parallel with other modules
• Movement times estimated using Fitts Law
• ddistance to move.
• wwidth of target
• k constant factor based on task type (e.g.,
  mouse movement)

28
Audition and Speech Modules

• Audition Module
• Audicon contains current auditory features
• Auditory features converted to declarative chunks
• Auditory input has temporal extent, so processing
  is often delayed
• Supports tones, digits, and simple strings
• No NLP
• Speech Module
• Simple generated speech output

29
Todays Lab

• Simple perceptual task
• See letter, give keyboard response
• Illustrates the operation of the visual and motor
  modules
• Lab assignment
• Extend demo code to handle harder perceptual task
• See three letters, return response for letter
  that is different from other two
• Many different solutions