Files
Colla_36961500_2020.pdf
Open access - Adobe PDF
- 7.27 MB
Details
- Supervisors
- Faculty
- Degree label
- Abstract
- This master thesis designs new efficient multi-agent algorithms that operate in open systems (i.e. systems subject to arrivals and departures of agents) and that aggregate the data held by each arriving agent for allowing them to estimate some external quantities in a decentralized way. Multi-Agent Systems are systems composed of independent, intelligent and interactive entities acting toward some common objective. Recognized for their advantages in solving large problems in a decentralized manner, their study generally considers a fixed composition of agents. However, arrivals and departures of agents may be unavoidable (such as computer failures in a network) and may impact the objective pursued by the agents. In that Open Multi-Agent Systems context, the agents may need to aggregate information from all the agents that have ever been in the system for estimating some external quantities but without using a growing memory. This is the case, for example, when each agent holds a measurement of a same random phenomenon for which each wants to estimate the distribution mean in a decentralized way. This latter problem is the focus of this master thesis. Its main challenge is to incorporate correctly the information from new agents without forgetting the information of the agents having left the system and without being too much impacted by noise. This master thesis establishes theoretical and empirical performance limitations for this problem in Open Multi-Agent Systems and presents new algorithms that solve it efficiently. The best designed algorithm reaches the performance limitations (within 1%) in most cases.