The Minimal Control Principle

1
The Minimal Control Principle

• Niels Taatgen
• Carnegie Mellon University

2
Problems with top-down control

• Brittle behavior
• Cannot handle unexpected events
• Cannot cope with missing knowledge
• Hard to account for multi-tasking
• Need for schedulers

3
Control in ACT-R
The contents of all the buffers together
determine which production will fire next
Declarative Module
Matching
Selection
Productions
Execution
Visual Module
Manual Module
External World
4
Top-down control
In most ACT-R models, the goal buffer largely
determines the next step to be taken
Declarative Module
Goal Buffer
Retrieval Buffer
Matching
Problem Buffer
Selection
Productions
Execution
Manual Buffer
Visual Buffer
Visual Module
Manual Module
External World
5
Bottom-up processing
However, other events can also satisfy production
conditions, for example a visual stimulus.
Declarative Module
Goal Buffer
Retrieval Buffer
Matching
Problem Buffer
Selection
Productions
Execution
Manual Buffer
Visual Buffer
Visual Module
Manual Module
External World
6
Combining top-down and bottom-up
Have as few as possible goal states in the goal
buffer, and have several rules that can react to
events
Declarative Module
Goal Buffer
Retrieval Buffer
Matching
Problem Buffer
Selection
Productions
Execution
Manual Buffer
Visual Buffer
Visual Module
Manual Module
External World
7
Example choice reaction task
goal done
goal retrieving
goal start
Wait for stimulus
Retrieve response
Check were done
Initiate response
Attend the stimulus
goal motor
goal attending
8
Brittleness example what if a stimulus appears
for which you have no response?
goal done
goal retrieving
goal start
Wait for stimulus
Retrieve response
Check were done
Initiate response
Attend the stimulus
goal motor
goal attending
9
No need for control states
Wait for stimulus
Retrieve response
Check were done
retrieval gt manual
visual gt retrieval
manual gt done
visual-location gt visual
Initiate response
Attend the stimulus
10
If the retrieval fails, the model will just pick
up the next stimulus
Wait for stimulus
Retrieve response
Check were done
retrieval gt manual
visual gt retrieval
manual gt done
visual-location gt visual
Initiate response
Attend the stimulus
11
Minimal Control Principle

• Have as few control states as possible
• In ACT-R the number of possible values for the
  goal buffer equals the number of control states
• The set of productions has to cover all possible
  combinations of buffer contents
• P control-states x buffer2 x ... x buffern
  --gt actions
• Number of rules you need linearly increases with
  the number of control states

12
Minimal Control Principle

• Assume that people strive for a knowledge set
  that has as few as possible control states
• Therefore in model construction create the model
  with the smallest possible set of control states
• In education and interface design strive for
  designs that encourage minimal control

13
Control in dual tasking
Attend Visual
Retrievewhich finger
Press thatfinger
Done
Attend Aural
Retrievewhich word
Say thatword
Can be ordered in 45 different ways, but the only
one that avoids all dual-task costs is
14
Attend Visual
Retrievewhich finger
Press thatfinger
Done
Attend Aural
Retrievewhich word
Say thatword
We can try to solve this by explicit reasoning
and planning, or
Rely on bottom-up processing to find the optimal
order,and not have any control states at all
15
How does it scale?the CMU-ASP task

• Subjects have to classify planes (tracks) on a
  radar-screen
• They have to do three things to classify a track
• Select one by clicking on it
• Use one of two classification methods, each of
  which is sometimes successful and sometimes not
• Enter the classification into the system

16
(No Transcript)
17
Representation ofInstructions
One control state for every instruction
Ideally, one control state for each set
of instructions Unit task
18
Instructions involve multiple steps

• Look at a track
• Find a track in peripherial vision
• Move attention to it
• Store the location

• Hook a track
• Move the hand to the mouse
• Move the mouse to the location of the track
• Click the mouse

Carrying out these steps in order is inefficient!
19
Novice
Encode Visual
Visual
Find location
Declarative
Click-object
Find-object
Find-object
Find-object
Click-object
Click-object
Mouse to object
Retrieve instruction
Retrieve instruction
Find-location
Retrieve Instruction
Find-object-attend
Retrieve instruction
Store object
Retrieve instruction
Hand-to-mouse
Retrieve instruction
Rules
Move handto mouse
Move mouseto location
Manual
Time
20
Production compilation learns rules that can fire
out of sequence
Click-object
Retrieve instruction
Hand-to-mouse
Compiled H-T-M
Move handto mouse
Move handto mouse
21
Encode Visual
Encode Visual
Encode Visual
Encode Visual
Visual
Find location
Find location
Find location
Find location
Find location
Move handto mouse
Location and hand
Attend plane
Store object
Attend plane
Compare new best
Attend plane
Compare old best
Move mouse plane
Attend plane
Compare new best
Move mouse plane
Attend plane
Rules
Move mouseto location
Move mouseto location
Manual
Time
22
Quantitative evidence
23
The minimal control principle and the type 1/type
2 distinction
Type 2 processes are independent of the control
state Can type 1 processes evolve into type 2
processes?
Declarative Module(Temporal/Hippocampus)
Intentional module(not identified)
Retrieval Buffer(VLPFC)
Goal Buffer(DLPFC)
Matching (Striatum)
Productions(Basal Ganglia)
Selection (Pallidum)
Execution (Thalamus)
Manual Buffer(Motor)
Visual Buffer(Parietal)
Visual Module(Occipital/Parietal)
Manual Module(Motor/Cerebellum)
External World
24
Take home message

• Try to minimize control in your models
• Try to minimize the need for control states in
  interface design
• Try to minimize the need for control states in
  designing instructions

25
Current Future work