Transcript Slide 1

Robot Recognition of
Complex Swarm Behaviors
Aisha Walcott-MAS622J-Dec. 11, 2006
Introduction
Dispersion
Orbit
A Swarm is a large collection of autonomous mobile robots
No centralized control
Group behaviors are produced from local interactions of
many individual robots
Goal is to develop a suite of primitive global behaviors that
combine to form more complex group programs
Courtesy James McLurkin
Project Goal
Build multi-classifiers to classify Complex Swarm Behaviors
3
3
3
3
2.5
2.5
2
2
1.5
1.5
1
1
0.5
2.5
2.5
2
2
1.5
1.5
1
1
0.5
0
0
0.5
1
1.5
2
2.5
Disperse
3
0.5
0
0
0.5
0.5
Orbit
1
1.5
2
0
0
2.5 0.5
Cluster
3
1
1.5
0
0
2
2.5
Example Features
0.5
3
Bubble Sort
1
1.5
2
2.5
hi source
source
source
low source
Approach
Collect raw behavior data sets
Determine Features (8D)
Pattern Recognition Algorithms
KNN
Neural Nets
Bayes Nets
Analyze results of each algorithm
KNN
Tested a range of values for nearest neighbors
random tie break
Overall Correct Classification
Average Class Classification
KNN for varying k on 4 Swarm robot behaviors
0.66
0.64
Cluster= 100%
Disperse = 12.5%
Clump = 50%
Orbit = 44%
Bubble Sort = 82%
percent correct
0.62
0.6
0.58
0.56
0.54
0.52
1
2
3
4
5
6
7
k neighbors
8
9
10
11
Neural Nets
Single Hidden Layer
Layer 1: nodes [50,70]
Max percent = 65%
Logsig
Two Hidden Layer
Layer 1: nodes [50,70]
Layer 2: nodes [25,25]
Max percent = 65%
1 Layer NN for with number of units = [50, 70] on 4 Swarm robot behaviors
2 Layer NN for with number of units = [50, 70] on 4 Swarm robot behaviors
0.8
0.65
0.7
0.6
0.55
percent correct
percent correct
0.6
0.5
0.4
0.5
0.45
0.4
0.3
0.35
0.2
0.1
0.3
0
5
10
15
number of nodes
20
25
0.25
0
5
10
15
number of nodes neighbors
20
25
Bayes Nets
Mapping to discrete domain by applying
k-means clustering to each feature
Preliminary Results
Cluster, Disperse,Clump,Orbit, Bubble Sort
1
2
8
Classification of Cluster
Possible bug in code
Modify the discrete mapping
Discussion
KNN and Neural Net performed well
Determining the mapping from real numbers
to a discrete domain may affect Bayes
Nets classifiers
Overall high classification of clustering
behavior
-Features tuned to behavior
-Not enough variety of samples
Need more samples of varying behavior
Next Steps
Feature selection-which group of features
work best for each classifier
Additional experiments to determine why
certain classifications are much better
Future
Use the temporal information to learn
hidden emergent sub-behaviors
Thank You