Machine Learning Approaches to

journeycartΤεχνίτη Νοημοσύνη και Ρομποτική

15 Οκτ 2013 (πριν από 3 χρόνια και 8 μήνες)

49 εμφανίσεις

Machine Learning Approaches to
Cognitive Parameter Acquisition

Terran Lane

University of New Mexico

terran@cs.unm.edu

Chris Forsythe, Patrick Xavier

Sandia National Labs

{jcforsy,pgxavie}@sandia.gov

Sandia’s Cognitive Modeling Framework


Computational models of human decision
-
makers


Models attention, perceptual cues, situational
awareness, decision making


Based on oscillatory models of activation


Spreading activation networks and feedback
loops between functional elements


Applications
--

data analysis, security, tutoring…


Bottleneck
: models hand
-
built/tuned


Expensive and slow!

The Big Picture

World

Cue
0

0

Cue
1

1

Cue
N

N

Situation
0

Situation
1

Situation
M

Actions/

Decisions


01


10


N1


NM



Automated Model Acquisition


High predictive accuracy


87% correct prediction of operator’s
interpretation of scenario (incl. relevance)


91% correct in recognizing situation only


Insights into operator decision
-
making process


Models are task & user specific


Only 26% overlap between users


Large effort in building and tuning models


Project goal: (semi
-
)automate acquisition of
parameters, network topologies, etc.


Prediction accuracy secondary concern


Roles for Machine Learning


Parameter acquisition


Interconnection weights


Activation levels


Oscillator frequencies


Network topologies


Inter
-
cue spreading activation network


Cue <
-
> situation relations


Feedbacks


Cues and situation identification

Parameter Acquisition

World

Cue
0

0

Cue
1

1

Cue
N

N

Situation
0

Situation
1

Situation
M

Actions/

Decisions


01


10


N1


NM



Parameter Acquisition: Issues


Superficially supervised learning


Observe features/cues and operator actions;
induce params (find


s.t. f

:C

A)


Similar to ANN backprop, EM, etc.


Many effective, well understood techniques


Problem: not just
high
-
likelihood

params


Actually want
params used by human operator


Much harder


observable stimuli don’t directly
reflect operator’s internal state


Cognitive plausibility constraint

Parameter Acquisition: Approaches


Additional instrumentation


Measure characteristics of operator


Biometrics


eye tracking, MEG, etc.


Expensive, not widespread


Maybe not informative to params anyway


Utility elicitation techniques


Software queries user about why decisions
were made / state of attention


Picks questions to maximally improve model


Emulates expert knowledge engineer

Network Topology Induction

World

Cue
0

0

Cue
1

1

Cue
N

N

Situation
0

Situation
1

Situation
M

Actions/

Decisions


01


10


N1


NM



Topology Induction: Issues


Find structure of interconnections between
variables (I.e., cues, situations)


Much

harder than parameter acquisition


Formally, maximum likelihood/MAP search
through all possible networks

Topology Induction: Issues


Find structure of interconnections between
variables (I.e., cues, situations)


Much

harder than parameter acquisition


Formally, maximum likelihood/MAP search
through all possible networks

L=137

Topology Induction: Issues


Find structure of interconnections between
variables (I.e., cues, situations)


Much

harder than parameter acquisition


Formally, maximum likelihood/MAP search
through all possible networks

L=137

L=238

Topology Induction: Issues


Find structure of interconnections between
variables (I.e., cues, situations)


Much

harder than parameter acquisition


Formally, maximum likelihood/MAP search
through all possible networks

L=137

L=238

L=493

Topology Induction: Issues


Find structure of interconnections between
variables (I.e., cues, situations)


Much

harder than parameter acquisition


Formally, maximum likelihood/MAP search
through all possible networks

L=137

L=238

L=493

L=318

Topology Induction: Issues


Find structure of interconnections between
variables (I.e., cues, situations)


Much

harder than parameter acquisition


Formally, maximum likelihood/MAP search
through all possible networks

L=137

L=238

L=493

L=318

Topology Induction: Approaches


Principles of structure search well understood


Gradient ascent, annealing, genetic search,
constrained search, etc.


Difficult in practice


Computationally intractable


Resulting models very sensitive to data


Spurious likelihood spikes


low confidence
models


Compounded by cognitive plausibility constraint


Can get leverage from cognitive plausibility,
though

Cue and Situation Identification

World

Cue
0

0

Cue
1

1

Cue
N

N

Situation
0

Situation
1

Situation
M

Actions/

Decisions


01


10


N1


NM



Cue and Situation Identification: Issues


Discern cues and whole environmental
situations employed by user


Related to constructive feature induction,
nonlinear projection identification, relational
learning, etc.


Search across all possible nodes/relations

N=2

N=3

Cue and Situations: Approaches


Cutting
-
edge ML problem


Direct elicitation is probably most promising
approach


Formulating search space/uncertainty reduction
not straightforward


Even user interface is difficult (naming synthetic
nodes/relations)

Conclusions


Decrease time/effort/cost to construct and tune
cognitive model


Constrained to correspond to human’s internal
model


Both bane and boon to automated model
construction


Insights into operator’s mental state/decision
-
making process


Requires/drives novel ML algorithms


Future work: all of it…

Questions?