Communication approaches in Multi-Agent Reinforcement Learning

Vladyslav Nechai

AutorIn

Vladyslav Nechai

Titel

Communication approaches in Multi-Agent Reinforcement Learning

Zitierfähige Url:

https://nbn-resolving.org/urn:nbn:de:bsz:14-qucosa2-937467

Erstveröffentlichung

2024

Datum der Einreichung

15.03.2024

Datum der Verteidigung

03.04.2024

Abstract (EN)

In decentralised multi-agent reinforcement learning communication can be used as a measure to increase coordination among the agents. At the same time, the essence of message exchange and its contribution to successful goal achievement can only be established with the domain knowledge of a given environment. This thesis focuses on understanding the impact of communication on a decentralised multi-agent system. To achieve this, communication is employed and studied in the context of Urban Air Mobility, in particular- to the vertiport terminal area control problem. A proposed in this work experimental framework, that promotes different information exchange protocols, allows to investigate if and how the agents leverage their communication capabilities. Acquired simulation results show that in the terminal area of a vertiport the aircrafts, controlled in a decentralised way, are capable of proper self-organisation, similar to the structured technique formulated in [Bertram and Wei(2020)]. A study of their communication mechanisms indicates that through different protocols the agents learn to signal their intent to enter a vertiport regardless of environment settings.

Freie Schlagwörter (EN)

multi-agent reinforcement learning, communication, urban air mobility

Klassifikation (DDC)

380

Klassifikation (RVK)

ZO 7015

AK 54540

GutachterIn