PhD Summary
CB_SeC (Context Based Service Discovery and Composition) Starting date:
April 1st 2002
Affiliation: PAI group, DIUF CS department, Fribourg University, Switzerland. Grant #20-65301Swiss National Science Fund (SNF), Welcome Project Supervisor: Pr Beat Hirsbrunner. Description: Pervasive computing is about building applications to bring computation into the real, physical world. The high degree of dynamism and heterogeneity of the resources involved in such applications makes service adaptation and interoperability a difcult task. This projects presents CB_SeC, a service framework that faces the adaptability problem by using context information in order to provide users with more tailored services in pervasive environments. Our approach promotes user’s context as well as services’s context. Results: coming soon :)) Keywords: Service discovery, service composition, ubiquitous computing, context-aware computing. Former Academic Projects
Collective Adaptation of a Heterogeneous Communicating Multi Robot System
Starting date:
November 1st 2002
Ending Date: March 2002 Affiliation: PAI group, DIUF CS department, Fribourg University, Switzerland. Grant #20-65301Swiss National Science Fund (SNF), Welcome Project Supervisors: Dr Michèle Courant and Dr Alessio Gaspar. Keywords: Collective robotics, emergence, dynamic communication, autonomous robots, Khepera. Description: This work
fits in the framework of autonomous agents. We are concerned with
collective phenomena (behaviors) and their issues, and more precisely the
way to carry out solutions that allow a heterogeneous communicating multi-agent
(multi-robot) system to adapt the collective behavior, in face of a changing
environment using a dynamic form of communication. This work is supported
by two types of experiments, namely those involving multi-agent simulations
and those involving real Khepera
robots. Our motivation of using robots to study multi agent systems (MAS) is twofold.
First we provide experimental set-up where agents are situated and
embodied. Results: A collective robotics framework, where a team of heterogeneous communicating mobile robots, operating without a supervisor and without a centralized control of their behaviors, adapt the collective behavior during runtime in face of a changing environment. Our results show that simplistic behavior rules implemented in a decentralized manner can lead to complex collective behaviors. The innovative aspect in our approach rests on a " system integrating communication as an active and dynamic component in the adaptation " , and not only as a static part of the robots interactions. The two main advantages of our approach are: Robustness and fault tolerance: robots can fail, or be removed from the collectivity to some extent without affecting the system. Scalability: our architecture allows a dynamic integration of robots. More robots can be added easily and immediately participate in the system, without writing new functionalities, as long as the connectivity in the communication network is maintained. Outlook: In the current formulation, the communication devices have to form a communication network over the execution time, because if a subset of robots wanders out of range, they would be isolated from the rest of the team. This could also be considered as a feature rather than a problem, because sometimes one might want to divide the team to solve tasks in different parts of the environment. However robots that have to cooperate to solve a task have to form a network most of the time. Till now we focused only on operational autonomy a next stage in the work will consist in studying behavioral autonomy, through learning approaches, in order to increase the adaptivity of the system. Another advantage will be possibly to make the collectivity reacts differently to different kind of events, just like some animals use a set of calls to signal different types of threats. Others: presentation at the workshop of the "Collective Intelligence specialized seminar" University of Fribourg, February 2002. Application of a Supervised Learning Method for Training of Neural Networks Starting date:
December 1st 1999
Affiliation: DI (Département d'Informatique), EPFL, Switzerland. Supervisors: Pr Wulfram Gerstner Keywords: Neural networks, supervised learning, training, DELVE database. Description: The goal of this project was practical implementation and experimentation with the neural networks using supervised learning method for training of neural network. We choose comp-activ database from the DELVE collection of databases. We used backprop algorithm for supervised learning with three methods of regularization : early stopping, weight decay and weight elimination. For all methods of regularization we first found the optimal parameters. For early stopping it is learning time, and for the two other methods that is parameter lambda. When we found optimal parameters we did training and testing of the neural network using these parameters and using division on training, testing and final patterns for which we got optimal parameters. Then we did retraining of the neural network using now training and testing data sets together for training. The third part, patterns for final performance measure we didn't touched until final performance measurement. Finally, we did final performance measurement and calculated final errors as it is done in DELVE database for methods they used. We compared our results with their results, and we compared results we got from our three methods. Results: Our experiment showed that if we do training for recommended number of training patterns (64, 128, 256, 512 and 1024) and we don't do retraining, final performance error goes down with increasing the number of data in training data set. Here with increasing the size of training data set error decreases slowly. On the other hand, after retraining error slightly goes down, it is constant, or it even goes up because of over-generalization. We can say that weight decay and early stopping have very similar results, while weight elimination is very bad for big sizes of training data sets (after retraining). Actually, all methods are good for small sizes of training datasets training, but when size of training dataset reach number of 256 decreasing of error after training is very slow, and after retraining we even have slight increasing. Virtual Reality in Psychiatry, Psychology and in Emergency Situations Starting date:
November 1999
Affiliation: LIG, (Laboratoire d'Informatique Graphique) DI (Département d'Informatique), EPFL, Switzerland. Keywords: Virtual Reality, psychiatry, psychology, emergency, simulations. Graphics Art: an Environment for Modeling and Illuminating Complex 3D Scenes Starting date:
November 1998
Ending date: November 1999 Affiliation: Vision, National Institute of Computer Science (INI), Algiers, Algeria. Supervisors: Dr Samy Ait Aoudia and Dr Amar Balla Keywords: Ray tracing, radiosity, image synthesis, global illumination models, . Description: The problem of light distribution, has been one of the main issues of realistic image synthesis during the last few years. The global lighting models can compute an accurate distribution of light within an environment. It envolves two methods : the radiosity which compute the diffuse part of the light, and the ray tracing which compute the specular part of the light. The goal of this work was a theoritical study of these two methods, the combination between them, and their practical implementation to different types of modelisation (CSG, Brep, surfaces ...etc) . Other ProjectsStarting date:
July 17th 2000
Ending date: March 2001 Affiliation: CAS1, Orthopedic and Biomechanics Division, Maurice E. Müller Institute for Biomechanics (MIB), Berne, Switzerland. Supervisors: Frank Langlotz Keywords: Collision, impingement, reconstruction, simulation. Description: The goal of this project was to implement an impingement detector for use in computer aided surgical interventions. Method: The algorithm is based on the combination of a lookup table and a linear transformation. It takes implicit object models from reconstruction of anatomical CT data which represent complicated anatomical structures. A look up table and linear transform are used to speed up the detection procedure. We do not assume any temporal and geometry coherences in order to meet the real-time demand. The method is of a general purpose in the sense that objects can be of any shape by assuming implicit solid models from reconstructions of CT data. Results: For any given transformation, the algorithm can perform impingement detection of two objects within 0.1 second. Others: Contribution with a team of three.
|