Transcript Document

I2RP/OPTIMA
Optimal Personal Interface by Man-Imitating Agents
Artificial intelligence & Cognitive Engineering Institute, University of Groningen, Grote Kruisstraat 2/1, 9712 TS Groningen, the Netherlands, http://www.ai.rug.nl
drs. Judith D.M. Grob (PhD student) dr. Niels A. Taatgen (supervisor) dr. Lambert Schomaker (promotor)
 Project Objective
 Future Plans
USER MODEL S
Sugar Factory Experiment
(Berry & Broadbent, 1984)
Task: Keep during two phases of
40 trials, the production P of a
simulated sugar factory at a target
value T, by allocating the right
number of workers W to the job.
Experimental Data
(Fum & Stocco, unpublished)
14
instances
12
Successes
Problem
•With software becoming more and more
complex, software design geared towards the
‘average user’ is insufficient, as different users
have different needs.
•Users differ in: goals, experience, interests,
knowledge.
•Possible Solution: Let the system maintain a
cognitive model of the user, which performs
the role of an intelligent agent that can inform
the interface on user-relevant adaptations.
 Current Work
10
3-3
9-9
3-9
9-3
8
6
4
compilation
through analogy
general
subconscious
rules
2
System Dynamics:
Pt = 2 W t - Pt-1 + Random Factor (-1/0/1)
0
phase1
phase2
Findings:
• Participants are better at reaching 3 than 9
• Implicit learning: participants improve but cannot verbalise knowledge
• Transfer: change of target doesn’t effect learning
declarative
conscious
rules
?
Two Computational Models
(in ACT-R)
Instance Model
Competing Strategies
(Taatgen & Wallach, 2002)
Objective
“To come to a methodology for the
development of adaptive user interfaces,
using the Cognitive Architecture ACT-R
(Anderson, 2002) as a modeling tool”
(Fum & Stocco, unpublished)
Model stores instances of
experiences with trials. It retrieves
these as examples to solve new
trials.
• Pro: Simple model
• Con: Cannot explain transfer
Model has 6 competing
strategies. The successful ones
are used more frequent over
time.
• Pro: Models all effects
• Con: Task-dependent strategies
Gain a better understanding of what happens
when people get more skilled at operating a
complex system, such as a software program.
References
Our Analogy Model
(in ACT-R)
• Contains simple, task independent analogy rules, which search for
common patterns e.g. repetition of values.
• Model applies analogy rules to instances retrieved from memory and
thus forms task-specific strategies to solve the task.
Three research phases:
Predictions by Analogy Model
Application
Agent
Application
Agent
Application
25
Agent controls
application
Agent
2.
3.
Agent learns
with user
User
20
Application
adapts, based
on agent
User
Successes
1.
15
10
5
Findings:
• Learning
• Difference between targets
3-3
But:
9-9
3-9
• No transfer
9-3
• Values are too high
0
phase1
Possible areas of adaptation:
•help function
•display of menu’s
•
•
•
•
Anderson, J. R. (2002). Spanning seven orders of magnitude: A
challenge for cognitive modeling. Cognitive Science, 26.
Berry, D.C., & Broadbent, D.E. (1984). On the relationship
between task performance and associated verbalizable
knowledge. The Quarterly Journal of Experimental Psychology,
36, 209-231
Fum, D. & Stocco, A. (unpublished). Instance vs. rule based
learning in controlling a dynamic system. Submitted to ICCM
2003.
Taatgen, N.A., & Wallach, D. (2002). Whether skill acquisition is
rule or instance based is determined by the structure of the task.
Cognitive Science Quarterly, 2, 163-204.
phase2
Next:
• Why doesn’t the model apply newly formed rules more often?
• Let model forget through decaying activation in memory
• Experiment with relative representations
634.000.002 (I2RP)