ACTR and Performance

1
ACT-R and Performance

• CS/ISYE/PSYC 7790
• Fall 2003

2
Administrivia

• Weekly ticket due today
• 1-page plan due in class on Wednesday

3
Warm-up question When good problem-solvers go
bad

• Mynatt kitchen example
• Making a recipe for chocolate chip cookies
• Phone rings mid-preparation
• Now, where were you?
• Whats going on here?
• Choir practice example
• Traveling to special choir rehearsal around 6 pm
• Turn onto highway going home instead of left
  towards rehearsal
• Why?
• (This is also used as an example of reactive
  planning)

4
Outline

• The problems of problem spaces
• How ACT-R chooses a production instantiation
• Creating a working memory
• Using timing to create a set of reasonable
  production instantiations

5
Claims

• Its not enough to have a set of reasonable
  production rules you also have to know when to
  use them
• This problem becomes harder, not easier, when we
  move from practical production systems to those
  used for cognitive modeling

6
The problems of problem spaces
Production rule instantiation
State

• A problem space
• All possible states of the world
• Transitions between states determined by
  production rules
• Example
• Subway a good example
• At Five Points
• North line to North Avenue
• No 13 bus to Tech
• At many points, multiple operators are available
• 1?(2 or 5)
• 2?(3 or 4)

7
The problems of problem spaces

• At state X
• Possible operators A, B, or C
• How do we choose?
• Put on your AI hat
• What solutions exist in AI?

8
Choosing the right production instantiation

• Well-bounded rules
• Just make sure that theres only one production
  rule that fits at any one time
• Distance metrics
• How do I choose between the bus and train at 5
  points?
• Selection heuristics
• Refractoriness, recency, specificity, strength
• Chunking (a la Soar) or composing sets of
  productions
• Relationship between this and case-based
  reasoning?

9
ACT-Rs model
Chunks
Productions
Possible instantiations
10
Data chunks in working memory

• Which data chunks do we want to consider?
• Each chunk has activation level (Ai)
• Most active chunks constitute working memory
• Activation Log odds of each data chunk being
  used in the current context
• Why log odds?
• Activation supplemented by related items
• Ai Bi sum(WjSji) for all j
• j items are those in goal or perceptual buffer
• Retrieval buffer too?

11
Why use log odds?

• Really interesting in probably, right?
• Probability ? 0..1
• Odds ? -infinity..infinity
• Log(probability) ? -infinity..0
• Log(odds) -gt -infinity..infinity, with nice
  midsection
• Can still compute the rough equivalent of
  probability using addition and subtraction
• P(AB) P(A)P(B)
• Log(P(AB))Log(P(A))Log(P(B))
• Doesnt quite work for odds, but close enough

12
ACT-Rs model
Chunks
Productions
Possible instantiations
13
Strength of productions

• Just like an activation level, in many ways
• Cover in more detail next time
• Production strength is learned over time
• Has a power effect

14
Which production instantiations to choose?

• Got set of active data chunks, strong
  productionsnow which are plausible?
• Keep in mind that there are essentially two parts
  to this
• Reasonable instantiations Set of instantiations
  that are reasonable, given limited prior data and
  activations
• Selected instantiation Instantiation that is
  chosen from that set
• Evaluation to determine the latter not the same
  as that which determines the former!
• How do we find the reasonable instantiations?

15
Finding reasonable instantiations using matching
time

• Novel twist to Andersons theory
• Reasonable instantiations a function of
• Level of activation
• Amount of time needed to instantiate the
  productions (in parallel)
• How do we calculate the amount of time needed to
  instantiate the productions, however?

16
Determining the instantiation time

• For each production instantiation, sum of clause
  instantiations
• I.e., clause instantiation is serial
• For each clause, instantiation time exponentially
  decreases with
• Activation level of data chunk for that clause
• Strength of that production
• TclauseBe-b(AiSp)
• Tpsum(Tclause) for all clauses
• Notice how the time calculation inverts the log
  relationship between an item and its chance to be
  used.

17
Instantiation times
p1
p2
p3
p4
p5
Begin
Fire
18
Instantiation times, with evaluation
13
p1
p2
25
p3
p4
10
20
p5
Begin
Fire
19
ACT-Rs model
Chunks
Productions
Possible instantiations
Fired production
20
Evaluation

• The evaluation is a function of
• Probability that it will lead to the goal
• Cost of the production
• Value of goal
• Like a distance metric, but more sophisticated
• Well see in the next chapter how this is learned
• Similar to some machine learning algorithms

21
Summary

• Selecting a production instantiation is a small
  problem in practical AI, a huge problem in
  cognitive modeling
• Assume very large number of productions and
  chunks
• For data chunks, we can use spreading activation
  to create a working memory
• Very tightly focused in ACT-R
• Think of kitchen example
• For production instantiations, we divide into two
  phases
• Picking set of reasonable instantiations based on
  timing considertations
• Timing in turn depends on data and production
  activations
• Evaluating set of reasonable instantiations to
  pick best instantiation

22
Next time Learning!

• Also remember to bring your short project
  description