Collaboration community formation in open systems for agents with multiple goals
Citation:GOLPAYEGANI, FATEMEH, Collaboration community formation in open systems for agents with multiple goals, Trinity College Dublin.School of Computer Science & Statistics.COMPUTER SYSTEMS, 2018
thesis.pdf (PhD thesis, examined and approved) 11.76Mb
Agents frequently coordinate their behaviour and collaborate to achieve a shared goal, share constrained resources, or accomplish a complex task that they cannot do alone. Forming an effective collaboration community in which agents are willing to cooperate, and have no conflict of interests, is the key to any successful collaborative process. Forming such communities has been addressed well in cooperative and closed multi-agent systems. However, it is particularly challenging in open multi-agent systems where agents are self-interested. Such agents are also likely to continuously and unpredictably leave and join the system and have multiple goals to pursue simultaneously. Existing research has addressed this challenge in open systems with utility-based or complementary-based approaches. Utility-based approaches focus on maximising self-interested agents' individual pay-off when sharing constrained resources. In complementary-based approaches, agents' individual skills are composed to accomplish a complex task or achieve a shared goal. In real-world applications, there are 2 main limitations with these approaches. First, it is impractical to assume that an agent will be exclusively either self-interested or cooperative. Agents cannot remain self-interested in open systems where resources are constrained and there is no central coordinator, because they will need to cooperate to sensibly share the resource. Second, it is too limiting to constrain agents, exclusively, to pursue either individual goals or shared goals during collaboration, as in real-world applications agents can simultaneously pursue multiple goals, including shared and individual goals. In open systems, agents need to identify the possible dependencies and conflicts between their individual goals when using constrained resources to pursue multiple goals simultaneously. Such dependencies affect agents' levels of self-interest and consequently their willingness to form collaboration communities. Given the circumstances, agents need a decentralised mechanism to acquire an understanding of other agents operating in their system, identify their goal dependencies, and adapt their level of self-interest to form effective collaboration communities. This thesis presents a fully decentralised approach to Collaboration Community FOrmation Model for agents with multiple goals in open systems (CCFOM). CCFOM presents a new social reasoning model and a new distributed community formation algorithm. CCFOM enables agents to pursue their individual and shared goals simultaneously in resource constrained open systems by forming effective collaboration communities. Each agent shares limited privacy-preserving domain knowledge and models its goal dependencies on other agents using social reasoning. The community formation algorithm is used to facilitate agents' decentralised decision-making process when forming collaboration communities. Using CCFOM, agents adapt their level of self-interest, and adjust their willingness to form effective collaboration communities. An application-independent simulator is used to evaluate CCFOM under varying levels of agents? mobility, and systems? density, and its application effectiveness is evaluated using smart grid and ride-sharing case studies. Evaluation metrics include measurements of collective and individual rates of success at achieving shared and individual goals. CCFOM is also evaluated against state of the art utility-based and complementary-based approaches to compare the communication and computational cost. Results show that CCFOM out-performs the utility-based and complementary-based approaches when agents pursue both individual and shared goals. It also decreases the computation and communication costs, when agents have multiple goals with varying dependency relations.
Author: GOLPAYEGANI, FATEMEH
Publisher:Trinity College Dublin. School of Computer Science & Statistics. Discipline of Computer Science
Type of material:Thesis
Availability:Full text available