Mojtaba Malek Akhlagh

Topic

Adaptive Resource Allocation in Multi-Agent Social Networks

Department of Computer Science

Date & location

• Tuesday, September 3, 2024
• 1:00 P.M.
• Virtual Defence

Examining Committee

Supervisory Committee

• Dr. Jens Weber, Department of Computer Science, University of Victoria (Supervisor)
• Dr. Kui Wu, Department of Computer Science, UVic (Member)
• Dr. Kin Fun Li, Department of Electrical and Computer Engineering, UVic (Outside Member)

External Examiner

• Dr. Glen Berseth, Department of Computer Science and Operations Research, Université de Montréal

Chair of Oral Examination

• Dr. Kenneth Stewart, Department of Economics, UVic

Abstract

Distributing resources among agents in social networks is an important and challenging problem. It involves deciding on assignment of a subset of resources to each agent based on the system objectives. Various instances of this problem can be observed in healthcare resource distribution, disaster management, cloud resource optimization, etc. In collaborative systems, different coordination techniques have been introduced in order to maximize overall social welfare. However, this problem becomes more complex in large-scale networks with limited connectivity among the agents. Moreover, in dynamic environments, where the set of tasks or resources change over time, an effective system needs to adapt to changes in the environment. Existing mechanisms fall short in addressing the social network constraints, and do not present efficient solutions when dealing with dynamic changes of supply and demand quantities. In this thesis, we view this social resource allocation problem (SRAP) as a multi-agent coordination problem. In a centralized approach, we consider a master agent with global knowledge, which makes decisions for all the agents. We present a greedy mechanism using an efficiency heuristic, and a learning-based mechanism by formulating the SRAP as a Markov Decision Process (MDP), and incorporating deep Q-learning. On the other hand, in a decentralized approach, we present a multi-agent protocol, which relies on local interactions among agents and their local knowledge only. The protocol enables the agents to negotiate with each other on allocation of resources to their tasks. It allows an agent in need of resources to concurrently negotiate with multiple providers and combine their resource contributions. We present greedy and learning-based mechanisms by integrating deep Q-learning into the negotiation process. In addition, the agents are able to cascade their corresponding information along the network, and apply timeouts in their messages. Hence, the decentralized protocol enables the multi-agent system to be self-organized, without relying on any central entity. We evaluate our approaches by developing simulation models of agents, tasks, and resources. We perform experiments on three main types of social networks, namely small-world, scale-free, and random networks. We conduct an empirical study of the performance of these approaches under varying conditions, such as resource availability, resource types, task requirements, etc. Our simulation results present a comprehensive analysis of various approaches across different types of social networks, by highlighting the strengths and limitations of centralized versus decentralized, as well as greedy versus learning-based approaches.