Leveraging nature-inspired techniques to semi-automate the design of multi-agent systems
Mc Donnell, Nicola
Mc Donnell, Nicola
Loading...
Publication Date
2023-07-11
Type
Thesis
Downloads
Citation
Abstract
Multi-agent systems (MAS) applications are used in many fields, from telecommu nications network management to traffic simulation. Their utilisation is increasing with emerging technologies such as Edge computing and the Internet of Things. This research aims to enhance the use of nature-inspired techniques that semi-automate the development of decentralised heuristics for resolving complex systems. It explores a series of frontiers in these areas focusing on different techniques and problem domains. The Gossip Contracts (GC) protocol is a novel communication protocol for multi agent systems to facilitate decentralised cooperation strategies. It is a generic protocol that can be applied to many problems. Hence, it was used to develop a novel de centralised dynamic virtual machine consolidation (DVMC) strategy. This strategy was compared to two well-known DVMC strategies, Sercon and ecoCloud, using a cloud data centre model derived from a real-world dataset. The assessment considered various cri teria and found that the GC-based DVMC had lower service level agreement violations than the other strategies. The Evolved Gossip Contracts (EGC) framework is a novel evolutionary computing based framework for designing decentralised heuristics. Using EGC, a decentralised heuristic for the NP-hard bin packing problem (BPP) was developed and shown to be better than two popular BPP heuristics, Best Fit and First Fit, for the datasets on which it was trained. Based on a literature review, EGC is the first evolutionary computing based framework evaluated on an NP-hard problem. The QD(λ) learning algorithm is a novel value-based MARL algorithm that leverages networked agents. It incorporates the advantages of transfer learning from the MARL QD learning algorithm and faster learning when rewards are delayed from the RL Q(λ) learning algorithm. It attained the lowest accumulative error for larger environments that require agents to transition through several states before receiving a reward. Thus, QD(λ) learning is suited to sizeable multi-agent systems learning a communication pro tocol, an emerging research area that has seen significant activity in the last few years.
Funder
Publisher
NUI Galway