dc.contributor.authorHou, Yaqing
dc.identifier.citationHou, Y. (2017). Evolutionary transfer learning for complex multi-agent reinforcement learning systems. Doctoral thesis, Nanyang Technological University, Singapore.
dc.description.abstractMulti-agent systems (MAS) are computerized systems composing of multiple interacting and autonomous agents within a common environment of interest for problem-solving. Since the behavioral strategies of agents in conventional MAS are usually manually defined in advance, the development of intelligent agents that are capable of adapting to the dynamic environment has attracted increasing attention in the past decades. In the past decades, reinforcement learning (RL) has been introduced into MAS as the learning paradigm of individual agents through trial-and-error interactions in a dynamic environment. Due to its generality and simplicity of use, the study of RL is rapidly expanding and a wide variety of approaches has already been proposed to exploit its benefits and applicabilities. A more recent machine learning paradigm of transfer learning (TL) has been proposed as an approach of leveraging valuable knowledge from related and well studied problem domains to enhance problem-solving in the target domains of interest. TL has been successfully used for enhancing RL tasks and many TL methodologies, such as instance or feature transfer, have been explored. Recently, the research on TL has been considered for enhancing multi-agent RL methods. In the context of computational intelligence, the science of memetics, especially memetic computation, has become an increasing topic of research. Many of the existing work in memetic computation has been established as an extension of the classical evolutionary algorithms where a meme is perceived as a form of individual learning procedure or local search operator in population based search algorithms. However, recent research shows memetic computation can become more meme-centric wherein memes transpire as units of domain information or knowledge building blocks useful for problem-solving. The intrinsic parallelism of natural evolution in such meme-centric computing may derive its strength from the simultaneous explorations of differing regions of a common problem domain and exploitative social interactions between the multiple agent learners. This dissertation hence seeks for the new study on a meme centric evolutionary knowledge transfer paradigm for problem-solving of multi-agent reinforcement learning systems. Specifically, this thesis presents an evolutionary transfer learning framework (eTL) which comprises a series of meme-inspired evolutionary knowledge representation and transfer mechanisms. The proposed framework serves to enhance the learning capabilities while addresses the limitations (e.g. blind reliance) of existing knowledge transfer frameworks. Subsequently, a novel transfer learning framework with predictive capabilities (eTL-P) is proposed to cope with the challenges arising in complex multi-agent systems where agents have differing or even competitive objectives. eTL-P endows agents with abilities to interact with competitive opponents, modeling their opponents, and hence predicting their behaviors accordingly. Further, to reduce the complexity of the opponent candidate models, a Top-K model selection method is proposed for selecting a smaller yet remarkably representative candidate model set from the entire model space. Last but not least, a summary of the future study on the evolutionary knowledge transfer paradigm is presented.en_US
dc.format.extent120 p.en_US
dc.subjectDRNTU::Engineering::Computer science and engineering::Computer applicationsen_US
dc.titleEvolutionary transfer learning for complex multi-agent reinforcement learning systemsen_US
dc.contributor.supervisorOng Yew Soonen_US
dc.contributor.schoolInterdisciplinary Graduate Schoolen_US
dc.description.degreeDOCTOR OF PHILOSOPHY (IGS)en_US
dc.contributor.organizationMulti-plAtform Game Innovation Centreen_US
dc.contributor.researchGame Laben_US

Files in this item


This item appears in the following Collection(s)

Show simple item record