Transcript Diapositiva 1 - Monica Reggiani

IEEE-RAS / IFRR SCHOOL OF
ROBOTICS SCINECE ON LEARNING
Lazise, Garda Lake, Italy, 24-28
September 2007
IEEE-RAS / IFRR SCHOOL OF
ROBOTICS SCIENCE ON LEARNING
GROUP 9
LUIS PAYALUCA LONINI
ALIREZA DERAKHSHAN
1. WHAT WE HAVE LEARNED.
2. NOVELTY DETECTION.
3. LEARNING WITH RECURRENT NEURAL NETWORK WITH
PARAMETRIC BIASES.
4. APPLICATIONS IN OUR RESEARCH AREA
1
IEEE-RAS / IFRR SCHOOL OF
ROBOTICS SCINECE ON LEARNING
Lazise, Garda Lake, Italy, 24-28
September 2007
1. WHAT WE HAVE LEARNED
• Use of a simple robotic platform to carry out experiments in
complex techniques of machine learning.
• We have dealt with simple external information - more
complex information should be added e.g. more sensory data.
• Learning by imitation
• Analytical models (system identification, policy learning by
imitation).
• Non Analytical models (learning with recurrent neural
networks with parametric biases).
• Statistical Analysis and Data Mining with Orange.
2
IEEE-RAS / IFRR SCHOOL OF
ROBOTICS SCINECE ON LEARNING
Lazise, Garda Lake, Italy, 24-28
September 2007
2. NOVELTY DET ECTION
• Working with readings from a Magellan’s 16 sonar sensors in a
wall following behavior.
• 1st train:
 s = 12 · (standard deviation)
 q = 0.6
• 58 kernels in the model base.
• Distances of each test data
to the nearest kernel of the
model base.
• Not a clear novelty among
the test data.
3
IEEE-RAS / IFRR SCHOOL OF
ROBOTICS SCINECE ON LEARNING
Lazise, Garda Lake, Italy, 24-28
September 2007
2. NOVELTY DETECTION
• 2nd train:
 s = 7 · (standard deviation)
 q = 0.6
• 321 kernels in the model base.
• Distances of each test
data to the nearest
kernel of the model base.
• Two possible candidates.
• The 2nd one (reading
100) is the novelty one
(maximum distance to
the nearest kernel).
4
IEEE-RAS / IFRR SCHOOL OF
ROBOTICS SCINECE ON LEARNING
Lazise, Garda Lake, Italy, 24-28
September 2007
3. LEARNING WITH RECURRENT NEURAL
NETWORK WITH PARAMETRIC BIASES.
• Building complex behaviors by combining simple primitive
behaviors.
• Each simple primitive can be coded with 2 biases.
Biases:
[0.68 0.40] Sinusoid
[0.73 0.36] Left
[0.19 0.78] Right
Biases: Keep Object Left
[0.99 0.0]
5
IEEE-RAS / IFRR SCHOOL OF
ROBOTICS SCINECE ON LEARNING
3. LEARNING WITH RECURRENT NEURAL
Lazise, Garda Lake, Italy, 24-28
September 2007
NETWORK WITH PARAMETRIC BIASES.
• Adding new primitives is possible
Biases: Obstacle Avoidance
[0.08 0.29]
QuickTime™ and a
Photo - JPEG decompressor
are needed to see this picture.
6
IEEE-RAS / IFRR SCHOOL OF
ROBOTICS SCINECE ON LEARNING
Lazise, Garda Lake, Italy, 24-28
September 2007
Applications in OUR Research Area
• Appearance-based Navigation
• These techniques can be applied to the localization and
navigation of a mobile robot using more complex
information (e.g. The information of the whole scene,
laser measures, etc.).
• It is necessary to analyze the scene and extract the most
relevant information.
• Classification of Playing Behavior
• Novelty Detection can be applied to categorize different
Playing Behavior based on some reference behaviors.
• Human motor learning models
• Machine learning techniques and Experiments with robots
can be useful to test hypothesis on neuroscientific theories
on how we do organize movements
• Novel control techniques can be applied to new generation
robots
7