Tasks, deliverables, and milestones
This task will investigate distributed control based on self-organizing systems, whereby decisions are taken by the units with no need of global information such as the number of the units, the characteristics of the heterogeneous units, the characteristics of the communication network. In between a completely centralized system and fully distributed, egalitarian organization, there is the possibility to organize different hierarchical architectures, which can themselves be built autonomously (through e.g. leader-election algorithms). The task will investigate the advantages and drawbacks of the different solutions, and provide indications as to the optimal trade-off for specific application requirements.
This task will investigate the theoretical bases and algorithmic implementation of methods to detect misbehaviors (caused by either faults or malicious attacks) of cooperating objects in a distributed system, and to recover through isolation and reconfiguration of the system.In distributed intrusion and fault detection systems, the detection and recovery tasks have themselves to be solved cooperatively by different agents, each possessing only a partial knowledge of the system status, by decisions made through negotiation, possibly based on trust or reputation. The task will also study security protocols that enable to dynamically check the admissibility of the units in the global system. For example, an observer/identifier can be supposed to be either a central decision maker, or to be implemented as a distributed task itself. In the latter case, security issues arising in dealing with potentially malicious agents propagating false information about neighbors (Byzantine behaviors) has to be taken into account. Robust decentralized algorithms to identify and isolate misbehaving nodes, based on partial observations by agents and on reputation building, are a difficult but crucial objective of this task, to enable building trusted systems of cooperating objects and agents.
Differently to pure sensor network applications, the distribution of data over time and space is a critical issue for networks of agents with actuation. The network capability of developing self-organizing and emergent behavior requires that information on the state as well as on the planned actions of the agents is available in parts of the network in which control actions are computed. Within this task, criteria for sufficient distribution of data with respect to time and space will be developed, considering relative and absolute coordinates for both quantities. Furthermore, algorithms will be deployed which achieve a proper synchronization of the agents in real-time based on a suitable data distribution scheme. Requirements arising from the synchronization of the agents determine the constraints for the control computation within the local regulation schemes of the agents, and control algorithms satisfying these constraints need to be developed.
Thermodynamics and conservations laws describe the macroscopic behavior of systems composed by large number of interacting objects (macroscopic and microscopic perspective). We can see these laws as a system description resulting form a model reduction of the original system. There is a need to obtain techniques for an “inverse model reduction” (inverse thermodynamics) that allow us to determine, from the reduced model, which are the laws at the microscopic level. This process of identifying the dynamics and rules obeyed by each agent has several important ramifications and applications. As an example, the issue of detecting how many and which different "species" of agents (as defined by their behavioral rules) are at work in a given environment has an evident bearing upon problems such as naturalistic observations, visual surveillance, intrusion detection, and traffic measurement.
In this task we will develop the local decision-making rules through cooperative learning so that the global system can adapt to the environment changes. To this aim we will develop adaptive control strategies for the coordination and control of multi-agent systems. In order to obtain an effective coordination and control strategy for multi-agent systems with distributed forms of reasoning, decision-making and action planning, we propose to use an approach based on adaptive learning, resulting in so-called multi-agent learning. The resulting system should be able to cope with unanticipated changes in the environment, to learn from past experience, to actively acquire and organize knowledge about the surrounding world, and to plan its future behavior. We will in particular consider situations in which the agents initially have a partial view of the world. Thereby, we will use and further develop unsupervised learning techniques (such as fuzzy clustering, reinforcement learning) and data-intensive methods (such as lazy learning). Furthermore, we also investigate how (cooperative) game theory can be used to obtain a structured design and analysis method for adaptive learning techniques. The resulting strategies should clearly guarantee robust performance under varying conditions.
In this task, the theoretical advances and the tools developed in tasks T4.1-T4.5 above will be applied to the concept studies considered in the proposal (as described in sections AD1, AD2, AD3):
- T4.6.1 Self-organizing systems and distributed coordination of vehicle-infrastructure (AD1a.2, AD1a.3, AD1b.1, AD1b.2)[CNRS, INRIA (lead), INRETS (lead), UKS, US, UNIVAQ, UNIPI, UNICA, TUDELFT , RUG, KTH. M1-M48]
- T.4.6.2 Self-organizing energy distribution systems (AD2)[INRIA, UNIVAQ, UNIPI, UNIPD (lead), RUG, KTH. M1-M48]
- T4.6.3 System architecture and distributed control of biological systems (AD3)[ETHZ, TUB, UNIVAQ, UNIPV (lead), TUDELFT, KTH. M1-M48]
The procedure will be cyclic, starting with an analysis of the case studies to extract the most relevant problems and requirements, informing a theoretical elaboration phase, followed by the application of the methods to models of the problems at hand (developed in the meantime in other Was). Indications extracted from the results will be used to modify the methods and tools, and so forth.
To read the available deliverables you must be a Level 1 Partner.
To upload a deliverable you must be a WP Leader.
Please log in to have access to deliverables and to download the template.