Discovering Optical Control Strategies: A Data-Mining Approach Romann M. Weber (
[email protected]) and Brett R. Fajen (
[email protected]) Cognitive Science Department Rensselaer Polytechnic Institute
17.86555
100
Experiment 1A, Ground Condition
Experiment 1B, Ground Condition
2 (176.0/7.0) 12
12
95
95
90
10
1 (sims) 10 (sims) 1 (subs) 10 (subs)
1 (436.0/3.0)
0.1
0.2
0.3
0.4
0.5
0.6
10
85
12
2 (35.0/2.0)
Does not impose a model on data but rather infers one from data.
1.0 0.8 0.6 0.4 0.0
2
4
6
8
10
Block
Experiment 1B, Air Condition
Multiple models can be "averaged" together using ensemble methods, allowing for a single model to describe collective subject strategies.
2 (121.0)
q
1 (31.0/13.0)
10
10
q
q
q
q
2
4
6
8
10
q
8
q
q
q q
6
q
q
While subject strategies can only be inferred, model strategies are transparent, allowing for new interpretations and experimental predictions.
q
q
8
q
q
6
q
inv.tau
q
1.409473
Block
As subjects become familiar with the task, single strategies (models) more accurately describe their data.
theta.dot
Strategies also become more consistent, as subject data can be modeled by previous blocks' strategies.
2 (173.0/5.0)
Decision trees make up only a small portion of the pattern-recognition and data-mining methods that can be applied to control-law research.
q
Block q
0.10795 1 (427.0)
Makes few initial assumptions about the data.
10
SUMMARY AND DISCUSSION
16
0.1
1 (407.0/1.0)
2
Block
8
8
inv.tau
Ideal Deceleration
6
6
Candidate control laws can be extracted from subject data by algorithms designed to identify patterns.
1 (sims) 10 (sims) 1 (subs) 10 (subs)
q
0.2
0.3
0.4
Radius
0.5
0.6
0.7
2 (193.0/15.0)
Block
0
Allows subject data to be analyzed collectively and individually.
4
4
Resulting models offer a highly accurate, compact, and intuitive representation of possible control strategies underlying actions.
0.066795
q
70
82
BENEFITS OF THE DATA-MINING APPROACH
2
14
1.459813
theta
q
q
12
Initial Speed
2.284893
4
84
q
2 (152.0)
Experiment 1A, Air Condition
2
q
10
inv.tau
1 (413.0)
0
q
1 (448.0/3.0)
Ideal Deceleration
q
8
12
90 88
q
2
inv.tau
80
q
6
Simulating Subject Data
Exp 1A Exp 1B
75
90
q
86
Mean Accuracy
q
Mean Prediction Accuracy
94 92
q
q
Initial speed variable, ground plane absent
8
4
0.7
1.815948
q
1 τ τ θ θ v τ
2 (190.0/1.0)
2 (164.0/3.0)
Trends in Accuracy Exp 1A Exp 1B
1 τ τ θ θ v τ
1 (sims) 10 (sims) 1 (subs) 10 (subs)
0
0
Models accurately and compactly capture patterns present in subject data.
1 (410.0/2.0)
Block
4
65
0.122046
Experiment 1B, Air Condition
q
q
Radius
Sign radius variable, ground plane absent
inv.tau
2
2
Block q
10
Block
Block
Condition
8
q
1.438791 theta.dot
6
6
8 6
q
q
4
Ideal Deceleration
70
q
q
65
70
q
4
Ideal Deceleration
85 75
80
Accuracy
q
Condition Initial speed variable, ground plane present
q
10
90 85 80 75
Accuracy
q
Air
4
v.over.tau
q
Ground
2
Experiment 1A, Air Condition
2 (142.0)
1 (424.0/7.0)
Air
10
Block
q
Ground
8
1.0
1 (458.0/2.0)
Experiment 1B
100
THREE POSSIBLE STRATEGIES (Fajen & Devaney, 2006)
6
1.929992
1.0
Experiment 1A
4
0.0
Model Accuracy (Classification Rates)
2
0.8
RESULTS
inv.tau
Variable−Call Probability
CONTROL LAWS IN BRAKING: THE STANDARD APPROACH
Sign radius variable, ground plane present
1.0 0.0
Data mining is a means of discovering these patterns.
0.2
The structure of the tree models can be interpreted as a representation of an underlying strategy.
0.6
A flowchart illustrating the taudot braking strategy (Lee, 1976).
0.4
Decrease Brake
0.8
Actions following control laws should embed patterns in subject data.
1 τ τ θ θ v τ
0.2
No
Tree models can be used to simulate subject behavior in novel conditions.
0.8
Decision trees are able to produce intuitive output representing patterns in data.
Increase Brake
Variable−Call Probability
ns
(optical info, actions)
1 τ τ θ θ v τ
0.6
Act io
Tree Model
Experiment 1B, Ground Condition
0.4
Environment
Yes
Is dτ/dt < -0.5?
(spatial info, actions)
Transformed Data
Experiment 1A, Ground Condition
0.2
a = π(s)
No
Raw Data
Patterns in the data that link optical information to actions can be extracted and expressed in decision-tree form (e.g. using C4.5/J48 tree-learning algorithm).
0.0
rm
Info
Successful performance of a task is a byproduct of acting to produce this pattern.
Maintain Brake
Trends in model structure reflect shifting strategies used by subjects.
Variable−Call Probability
n atio
Yes
Is dτ/dt = -0.5?
Control laws can be interpreted as algorithms for task control and expressed in an intuitive graphical form.
Spatial data can be transformed to recreate the optical space subjects are responding to.
RESULTS (Continued)
0.6
Visual strategies that prescribe actions that produce a target pattern of optic flow. Agent
MINING POSSIBLE CONTROL LAWS FROM DATA
Control Brake
0.4
THE NATURE OF CONTROL LAWS
0.2
http://panda.cogsci.rpi.edu
CO GN I TI VE SCI EN CE
Variable−Call Probability
PandA Labs
4
6
8
10
12
14
1 (sims) 10 (sims) 1 (subs) 10 (subs)
16
Initial Speed
(Tree models represent simplified ensemble summaries of individual subject models.)
Methods are capable of handling both discrete and continuous actions. These methods allow for control-law research into tasks with no obvious a priori strategies.
