The Research On Adaptive Task-Oriented Cooperation Mechanisms For Multi-Agent Systems In Unreliable Social Networks

The study of multiagent systems (MAS) is an important research topic in distributed artificial intelligence (DAI) domain. One of the key problems in MAS is the task-oriented multiagent cooperation which focuses on solving the problem of how to coordinate multiple agents to accomplish tasks efficiently that cannot be accomplished by an agent solely. This dissertation also focuses on this problem in unreliable social network environments introduced below.The research of social networks (SNs) has gained great interests during the recent years; because of the superiority of MAS in modeling the system consisted of autonomic entities, the research of MAS in SNs has also attracts much attention, which studies problems in MAS with typical characteristics of SNs. Nowadays, the unreliability has become one of the key characteristics of SNs because of the involvement of malicious agents which take malicious behaviors in task-oriented cooperation, and it will sigiuficantly reduce the reliability of the cooperation and the system performance (e.g. the higher costs or lower utilities). Besides, the properties of multiplexity and unidirectional interaction of SNs will also make the difficulty to guarantee the cooperation reliability and the system performance.By considering the involvement of malicious agents in task-oriented multiagent cooperation, the following adaptive mechanisms have been proposed from the task allocation perspective and the task execution perspective, respectively: · Adaptive task allocation mechanism optimizing unreliable resource accessIn multiagent systems, the accomplishment of a task can be modeled as the successful access to resources that are distributed in the system. If malicious agents are involved in the cooperation, the access to resources will be unreliable which may further cause the failure of that task. To solve this problem, the heterogeneity of resource reliability caused by the involvement of malicious agents is first considered. Then a negotiation strength-based reputation model and a corresponding reward/punishment approach have been proposed which can be used to assess the reliability of an agent and the resources it owns. Finally by considering the network constraint to the resource access and the reliability of agents together, the mechanism can adjust the allocation of tasks adaptively in order to improve the reliability of resource access and to reduce the time costs of the task accomplishment.·Adaptive task execution mechanism based on unreliable resource reorganizationDuring the task execution phase, agents can make reconfiguration to optimize the performance of task execution. Hence, in order to improve the reliability of task-oriented multiagent cooperation and further the system performance in unreliable social networks, an agent division and fusion based adaptive task execution mechanism has been designed. By employing this mechanism, agents are able to achieve unreliable resource reorganization of its local environment in a learning protocol. Accordingly, reliable resources in the system can be clustered together in some local areas after some tasks’execution. Hence, the probability of tasks accessing unreliable resources can be reduced in such areas, which finally improve the reliability of task execution and the system utility.More importantly, the above two mechanisms can be aggregated together to achieve a better performance in task-oriented multiagent cooperation in unrealible SNs.Then, by considering the influence of the topology characteristics of unreliable SNs on the performance of task-oriented multiagent cooperation, i.e., the multiplexity of network and the property of unidirectional interaction, the following mechanisms have been proposed respectively:·Adaptive task allocation mechanism compatible with unreliable multiplex social networksThe multiplex of networks is an emerging characteristic of unreliable social network. In such unreliable multiplex social networks, the reliability of task-oriented cooperation and the system performance (time costs or utility) are influenced not only by the heterogeneity of resource reliability caused by the involvement of malicious agents, but also by the heterogeneity of resource access costs and reliability in each network layer. Therefore, an integrated reputation model and a corresponding adaptive approach is firstly proposed, which can assess the reliability of agents and network layers respectively. Then, by considering the resource access reliability and the network constraint to the resource access caused by the multiplex network structure, the network layer-oriented adaptive task allocation mechanism and agent-oriented adaptive task allocation mechanism have been proposed to achieve the objectives of optimizing the reliability of resource access and the system performance, respectively. Employing these two mechanisms, the allocation of tasks can be adapted to achieve the optimizing objectives introduced above.·Adaptive task transfer mechanism considering unreliable unidirectional interactionThe property of unidirectional interaction is another emerging characteristic of unreliable social network. In networks with this characteristic, agents will transfer the task in order to jointly accomplish tasks due to the lack of bidirectional interactions between agents. However, the involvement of malicious agents may cause the reduction of system performance of task transfer, and more seriously, the assessment of reliability of agents are hard to conducted due to the lack of bidirectional interactions. To solve these problems, a novel reputation detection pattern "From path to individual" has been firstly proposed. Evidence space of path reputation can be constructed by the proposed detection approaches through the task transfer history, and then the reliability assessment of agents can be performed through the evidence space transformation. Based on this reputation model, the adaptive task transfer mechanism compatible with unreliable unidirectional interaction is devised, which can adjust the task transfer path adaptively so that the number of malicious behaviors in the system can be largely reduced and finally the objective of system performance optimization is satisfied.Overall, this thesis deeply studies the problem of task-oriented cooperation in multi-agent systems in unreliable social networks, and a series of adaptive mechanisms have been proposed, which can optimize the system performance by considering the key properties of the unreliable social networks.