Three Misconceptions about ACTR

1
Three Misconceptions about ACT-R

• Niels Taatgen
• Carnegie Mellon University

2
Misconception 1
ACT-R is rational
3
Rationalism

• In principle, all knowledge including
  scientific knowledge can be gained through the
  use of reason alone

4
ACT-R has more in common with Empiricism

• All knowledge comes to us through the senses and
  through experience
• (or, by encoding it in declarative memory)

5
What about the R in ACT-R?

• The architecture doesnt ACT Rationally
• But is constructed rationally (by evolution)
• For example, declarative memory is a rational
  solution to storing memories of different
  importance in an effective way

6
Misconception 1
ACT-R is rational
7
Misconception 2
ACT-R is a rule-based system
8
Properties of Rules

• Rules represent general truths about the world
• If p-gtq and p then q
• All birds can fly
• When the light is red you should stop
• A rule matches an arbitrary set of conditions (or
  premisses), and then takes a set of actions or
  reaches a set of conclusions
• Rules are often used as basic elements of
  reasoning, either forward chaining towards to the
  goal or backward chaining from the goal
• None of this is really true for ACT-R

9
Overview ACT-R in a Diagram
Declarative Module(Temporal/Hippocampus)
Intentional module(not identified)
ACT-R Cycle Matching Production rules that
match the current contents of the buffers are
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
10
Overview ACT-R in a Diagram
Declarative Module(Temporal/Hippocampus)
Intentional module(not identified)
ACT-R Cycle Selection Select the production
rule with the highest Utility
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
11
Overview ACT-R in a Diagram
Declarative Module(Temporal/Hippocampus)
Intentional module(not identified)
ACT-R Cycle Execution The selected production
rule modifies contents of buffers Modules operate
asynchronously from central cognition
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
12
Example model

• Count from 2 to 5

13
The counting model
3 then 4
2 then 3
4 then 5
Declarative Module(Temporal/Hippocampus)
StartCount from2 to 5
Retrieval Buffer(VLPFC)
Goal Buffer(DLPFC)
R1 request first count-fact
Productions(Basal Ganglia)
R2 say current request next
R3 say last done
Vocal Buffer(Motor)
Visual Buffer(Parietal)
Visual Module(Occipital/Parietal)
Vocal Module(Motor/Cerebellum)
Time 0
External World
14
The counting model
First rule matches the fact that we are starting
the count
3 then 4
2 then 3
4 then 5
Declarative Module(Temporal/Hippocampus)
StartCount from2 to 5
Retrieval Buffer(VLPFC)
Goal Buffer(DLPFC)
R1 request first count-fact
Productions(Basal Ganglia)
R2 say current request next
R3 say last done
Vocal Buffer(Motor)
Visual Buffer(Parietal)
Visual Module(Occipital/Parietal)
Vocal Module(Motor/Cerebellum)
Time 25 ms
External World
15
The counting model
R1 has two actions request the successor of (in
this case) 2, and note in the goal that we are
counting
3 then 4
2 then 3
4 then 5
Declarative Module(Temporal/Hippocampus)
Countingfrom2 to 5
Retrieval Buffer(VLPFC)
Goal Buffer(DLPFC)
2 then ?
R1 request first count-fact
Productions(Basal Ganglia)
R2 say current request next
R3 say last done
Vocal Buffer(Motor)
Visual Buffer(Parietal)
Visual Module(Occipital/Parietal)
Vocal Module(Motor/Cerebellum)
Time 50 ms
External World
16
The counting model
The declarative system now fulfills the request
3 then 4
2 then 3
4 then 5
Declarative Module(Temporal/Hippocampus)
Countingfrom2 to 5
Retrieval Buffer(VLPFC)
Goal Buffer(DLPFC)
2 then ?
2 then 3
R1 request first count-fact
Productions(Basal Ganglia)
R2 say current request next
R3 say last done
Vocal Buffer(Motor)
Visual Buffer(Parietal)
Visual Module(Occipital/Parietal)
Vocal Module(Motor/Cerebellum)
Time 150 ms
External World
17
The counting model
The retrieved fact and the goal now fulfill the
conditions of R2
3 then 4
2 then 3
4 then 5
Declarative Module(Temporal/Hippocampus)
Countingfrom2 to 5
Retrieval Buffer(VLPFC)
Goal Buffer(DLPFC)
2 then 3
R1 request first count-fact
Productions(Basal Ganglia)
R2 say current request next
R3 say last done
Vocal Buffer(Motor)
Visual Buffer(Parietal)
Visual Module(Occipital/Parietal)
Vocal Module(Motor/Cerebellum)
Time 175 ms
External World
18
The counting model
R2 will do several things at the same
time Increase the counter in the goal Request
the next number Initiate saying the number
3 then 4
2 then 3
4 then 5
Declarative Module(Temporal/Hippocampus)
Countingfrom3 to 5
Retrieval Buffer(VLPFC)
Goal Buffer(DLPFC)
3 then ?
R1 request first count-fact
Productions(Basal Ganglia)
R2 say current request next
R3 say last done
Say 2
Vocal Buffer(Motor)
Visual Buffer(Parietal)
Visual Module(Occipital/Parietal)
Vocal Module(Motor/Cerebellum)
Time 200 ms
External World
19
The counting model
Now again several things will happen in
parallel Declarative memory tries to find the
number after 3 The Vocal system starts saying
two R2 can only fire again when both are done
3 then 4
2 then 3
4 then 5
Declarative Module(Temporal/Hippocampus)
Countingfrom3 to 5
Retrieval Buffer(VLPFC)
Goal Buffer(DLPFC)
3 then ?
3 then 4
R1 request first count-fact
Productions(Basal Ganglia)
R2 say current request next
R3 say last done
Say 2
Vocal Buffer(Motor)
Visual Buffer(Parietal)
Visual Module(Occipital/Parietal)
Vocal Module(Motor/Cerebellum)
Say 2
Two
External World
Time 600 ms
20
The counting model
Several cycles we have finished counting, so R3
fires and says the last number
3 then 4
2 then 3
4 then 5
Declarative Module(Temporal/Hippocampus)
Countingfrom5 to 5
Retrieval Buffer(VLPFC)
Goal Buffer(DLPFC)
3 then ?
5 then ?
R1 request first count-fact
Productions(Basal Ganglia)
R2 say current request next
R3 say last done
Say 5
Vocal Buffer(Motor)
Visual Buffer(Parietal)
Visual Module(Occipital/Parietal)
Vocal Module(Motor/Cerebellum)
Time 1650 ms
Four
External World
21
Productions in ACT-R

• Traditionally they are called rules
• But they are much more primitive than rules in
  other systems
• So calling them rules is only confusing people
  who already have some conception of what a rule
  is
• In the past ACT-Rs working memory elements
  have been relabeled to chunks. Maybe it is time
  to drop the term rule?

22
Misconception 2
ACT-R is a rule-based system
23
Misconception 3
ACT-R has a central bottleneck
24
ACT-R Diagram
Declarative Module(Temporal/Hippocampus)
Intentional module(not identified)

Retrieval Buffer(VLPFC)
Goal Buffer(DLPFC)
Matching (Striatum)
Why isnt this the central bottleneck?
Productions(Basal Ganglia)
Selection (Pallidum)
Execution (Thalamus)
Manual Buffer(Motor)
Visual Buffer(Parietal)
Visual Module(Occipital/Parietal)
Manual Module(Motor/Cerebellum)
External World
25
ACT-R Diagram

• Any of the subsystems can be the bottleneck
• Often declarative memory is, or the visual system

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
26
Productions in the middle are the least likely
bottleneck

• Cycle is only 50 ms
• Learning (production compilation) can overcome
  any bottleneck in the center

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
27
Production compilation combines two rules into one

• Any two rules that fire in sequence are
  eventually combined into one
• In the limit, average central bottleneck is never
  longer than 25 ms

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
28
The true bottleneck is often Declarative memory

• For example, press key when the letter is a
  vowel
• Production compilation cuts our declarative
  retrievals, learning rules like
• If you see an E press the key

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
29
Misconception 3
ACT-R has a central bottleneck
30
Type 1 vs. Type 2 theories
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
31
One of my questions for this workshop
To what extend are type 1 and type 2 theories
different? Example Peterson Simon Subitizing
experiment
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