Download Support Vector Machines - Robotics Group of Bogazici University

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

AI winter wikipedia , lookup

Artificial intelligence in video games wikipedia , lookup

Technological singularity wikipedia , lookup

Embodied cognitive science wikipedia , lookup

Philosophy of artificial intelligence wikipedia , lookup

History of artificial intelligence wikipedia , lookup

Ethics of artificial intelligence wikipedia , lookup

Intelligence explosion wikipedia , lookup

Existential risk from artificial general intelligence wikipedia , lookup

Transcript
Artificial Intelligence Lab.
Techniques for Improving Vision and
Locomotion on the Sony AIBO Robot
by Quinlan M., Chalup S., Middleton R.
E. Itır Karaç
Artificial Intelligence Laboratory
Department of Computer Engineering
Boğaziçi University
Boğaziçi University
Artificial Intelligence Lab.
Outline
• Introduction
• Hardware Environment
• Techniques and Tasks
– Color detection using SVMs
– Collusion detection using SVMs
• Conclusion
Bogazici University
2
Artificial Intelligence Lab.
Introduction
Bogazici University
3
Artificial Intelligence Lab.
Hardware and Environment
• Sony AIBO entertainment models ERS-210 or ERS-210A
– 64-bit RISC processor with clock speed 192 MHz and 384 MHz
– programmed in C++ using the Sony’s OPEN-R environment
– the use of servos gives the robot 20 degrees of freedom
• RoboCup Legged League
Bogazici University
4
Artificial Intelligence Lab.
Techniques
• Support Vector Machines
• Multi-class SVMs
• One-class SVMs
Bogazici University
5
Artificial Intelligence Lab.
One-Class SVMs
• Idea: try to find a sphere with minimum volume, containing most of the
data objects
Bogazici University
6
Artificial Intelligence Lab.
Formulation of One-class SVM
• describe the sphere with center a and radius R.
min F ( R, a,  i )  R 2  C i  i
s.t. ( xi  a )T ( xi  a )  R 2   i
i,  i  0
• The center of the sphere is a linear combination of some of the data
objects, called support objects.
• Support objects and corresponding weights are obtained by solving this
optimization problem
Bogazici University
7
Artificial Intelligence Lab.
Tasks
• Vision system for most teams consists of 4 main tasks:
–
–
–
–
Color Classification
Run Length Encoding
Blob Formation
Object Recognition
Bogazici University
8
Artificial Intelligence Lab.
Color Classification
• Color Classification Task
– Images are taken from the camera in YUV bitmap format
– Each pixel in the image is assigned a color label using a lookup table.
• Initial generation of the LUT is critical and a new LUT has to
be generated with any change in the lighting condition.
• Currently this is done manually by taking hundred of images
and assigning a color label pixel-by-pixel-basis
• This process is time consuming and may still contain holes and
classification errors
Bogazici University
9
Artificial Intelligence Lab.
Method by Shapiro & Stockman
• Convert existing LUT values from YUV to the HSI color space
• Fit an ellipsoid E, which can be represented by the quadratic form:


E ( x 0 , Q)  x  R 3 : ( x  x0 )T Q 1 ( x  x0 ) 1 where x 0 is the center of the ellipsoid
• This problem is linear in the unknowns and leads to the convex
optimization problem
• This formulation tries to find the ellipsoid such that the sum of the squares
of the lengths of the principle axis is minimum
• Disadvantage:
– restricting the shape of possible regions
– duplicates and potential outliers should be removed manually before the ellipsoid is fitted
• Advantage: a simple representation
Bogazici University
10
Artificial Intelligence Lab.
Proposed Method
• An individual one-class SVM is created for each color.
• With an extremely low υ, and large γ the boundary formed by
the desicion function contains (1- υ) of the training points
• Advantage: SVM simultaneously removes the outliers
• SVM can be used in to situations
– Set up phase at a competion
– Updating an existing LUT
Bogazici University
11
Artificial Intelligence Lab.
Results
Bogazici University
12
Artificial Intelligence Lab.
Results
Bogazici University
13
Artificial Intelligence Lab.
Results
Bogazici University
14
Artificial Intelligence Lab.
Another Task: Collusion Detection
• Previously statistical methods are used
– Requires 6MB of memory
– It relies on domain knowledge
– Extremely low computational expense
• One-class SVM is employed as a novelty detection mechanism
• SVM decision function will return +1 for normal step, -1 for
vaulty steps
• Aim: minimize falkse positives
Bogazici University
15
Artificial Intelligence Lab.
Questions ?
Bogazici University
16