Grant: Discovery Grants Program - Individual

The research program focuses on developing innovative multi-agent models and algorithms, emphasizing agents' abilities to reason about their environments and decisions while considering dynamic trust and verification of complex systems. Trust plays a critical role in decision-making processes and system behaviour. Verification ensures the correctness of the system design and its dynamic behaviour, which increases the system's trust. The program aims to address these challenges by incorporating logic and reasoning-based multi-agent systems with emerging Deep Multi-Agent Reinforcement Learning (DMARL), Federated Reinforcement Learning (FRL), and Multi-Expert Demonstration Cloning (MEDC) techniques. The program has the following objectives: 1- Logic-based Dynamic Trust Modelling: The program will introduce reasoning-based multi-agent models that utilize formal and computational logics to capture the complexity of interactions among agents and dynamic trust. This involves adapting logical frameworks to multi-agent dynamic systems with uncertainty and inconsistency considerations. 2- Trust-Aware Deep Reinforcement Learning: The proposal aims to integrate formal dynamic trust modelling into DMARL, FRL, and MEDC algorithms, enabling agents to adapt their behaviours based on evolving trust relationships. Trust-aware multi-agent reinforcement learning enhances cooperation and collective decision-making. 3- Deep Learning-based Formal Verification: New formal verification through model checking procedures will be introduced for multi-agent dynamic trust logics. This includes symbolic and transformation-based model checking algorithms with soundness and computational complexity analysis. Deep learning techniques will be explored to improve the efficiency and scalability of model checking. 4- Application of the developed frameworks: The developed models, algorithms, and frameworks will be applied to address real-world challenges in various domains such as robot swarms, healthcare applications, and the Internet of Things/Internet of Vehicles (IoT/IoV). To tackle the challenges of limited real-world interaction data, the program will explore offline reinforcement learning methods, which can build models with fewer interactions. Various libraries, tools, and real-world datasets relevant to each application will be utilized. The research team will continue the practice of making implementations open source and reproducible, contributing to the enhancement of available machine learning libraries. The identified objectives present novel challenges yet to be tackled. The expected significance lies in advancing multi-agent systems, trust modelling, and model checking. The research will have applications in robotics, autonomous systems, cybersecurity, and complex network analysis, leading to more reliable and secure systems. It addresses novel challenges, offering more realistic representations of trust dynamics and efficient verification of complex systems.