Meritxell Vinyals

e-mail: meritxell.vinyals@gmail.com

About my research

My name is Meritxell Vinyals and I am a PhD student at the Artificial Intelligence Research Institute (IIIA-CSIC) in Barcelona, Spain. My work is currently focused on designing reconfiguration and self-organization mechanisms for multi-agent systems that allow agents to coordinate their activities in a decentralized and autonomic way. Concretely I am working to apply such mechanisms to coordinate sensor networks, networks composed of a large number of sensors distributed over an area with different computation, communication and sensing capacities. I hope that the Complex Systems Theory helps me to understand and regulate the global system behavior that emerge from local agent interaccions in multi-agent systems. I am really excited about this course and to meet all of you in Santa Fe! I think that meeting such interesting people coming all from studying very different problems will be an incredible experience!

A little more about me

In my free time I like traveling, reading, going outdoors (to the montain or to the beach, it doesn't matter) or simply having a drink with some friends!

Answers to Dan RockMore questions

• What are your main interests?

My main interests lie on distributed artificial intelligence (DAI) and concretely on designing coordination mechanisms for distributed multi-agent systems (MAS). As a practical application, I am looking to apply and adapt such mechanisms at the domain of sensor networks.

• What sorts of expertise can you bring to the group?

I hold a degree on Computer Science so I am used to programming. The languages in which I have more experiencie are: Java, C, C++, Matlab and php. I have also experience on running and program agent-based simulations, mainly using REPAST (http://repast.sourceforge.net/) platform. From my research I got theoretical knowledge in fields like Computational Mechanism Design, Market mechanisms and Multi-agent Systems. Currently my research is focused on pursuing computational-efficient distributed coordination mechanisms inspired on distributed graphical models and their inference algorithms. Thus, I have acquired some background on graphical models and message-passing algorithms used to solve them. Moreover, last year during my Phd courses I took classes on search algorithms, machine learning and logics.

• What do you hope to get out of the CSSS?

First thing I want to get out from CSSS is to get some understanding about how to model and deal with interactions in a system in order to apply and migrate some ideas to the coordination mechanisms in multi-agent systems. Usually, the complexity of how to design a multi-agent system emerge from the interactions and dependencies among the agents, not from the agent complexity itself. I think that complex system community have been years studying the same problem from different points of views. Furthermore I am interested on learning more about concepts and theories on emergence, self-organization and bio-inspired solutions. Finally I hope to have the opportunity to meet several classmates with whom I may have similar or complementary problems and share some ideas with them.

• Do you have any possible projects in mind for the CSSS?

Currently I have several aspects related to my research that I think are quite connected with complex systems. Here are my three proposals:

. Nearly-descomposable systems. Nearly-descomposable systems are systems in which although there are a lot of interactions and dependencies among the components of the system, most of these interactions nearly not influence the global system behavior. In general we can model these interactions as weak interactions but we can not obviate them. However from a computational point of view, sometimes to find the optimal solution for such systems is too hard and researchers have to work with approximate algorithms. In such context, the next question emerge: could the structure of these nearly-descomposable system be exploited in order to cut only weak interactions and reduce the computational complexity of the problem whereas the quality of the solution is only slightly affected? A project that tries to find methods in order to identify such interactions to reduce such complexity would be ambitious but also great!

. Transfer learning. Learn in dynamic environments has the problem that the learning context changes continuously over time and agents may need to start learning from scratch. For example if the learning context for one agent is defined over the interactions with its neighborhood , if neighbors change it has to throw away all the experience accumulated so far. Transfer learning (http://en.wikipedia.org/wiki/Inductive_transfer , http://en.wikipedia.org/wiki/Transfer_of_learning) mechanisms allow to transmit what agents have learned so far in a former learning context to some new, similar context. Hence one project could be to to apply and analyze the effect of using transfer learning mechanisms in some dynamical system.

. Lead with uncertainty. It is very common that in real world domains agents have uncertainty about the context of their decisions which is infered through communication and observations. This problem is very complex in a centralized way and gets more complex in decentralized environments where each agent takes each own decision. In artificial intelligence these systems are modeled as DEC-POMDPS. So study some application of DEC-POMDps in some domain or the relationship of such model with other similar models used in other fields could be also an interesting project!

These are the ideas that came to my mind when I was thinking on complex systems and multi-agent systems. Anyway I am up for working in other proposals, to explore other lines of connection between complex systems and MAS.