Multi-access edge computing (MEC) combines the elements of telecommunications networking and information technology to provide cloud computing services at the edge of the network. Different from the traditional cloud computing system, MEC is usually implemented at cellular base stations or other edge nodes. This position of MEC servers allows the execution of applications within close proximity of end-users, which substantially reduces end-to-end delay, congestion, and burden on the backhaul. However, the efficiency of the MEC system is greatly affected by the task offloading policy (since there can be multiple computation nodes, MEC servers, and users, and tasks can have diverse characteristics) and computation resource allocation decisions (computation resources are limited, and a large number of users affect the task execution delay). Therefore, we formulate the problems as optimization problems and design algorithms using techniques such as heuristics and machine learning (Deep Reinforcement Learning) to find an effective task offloading and resource allocation policy that meets different requirements of tasks, users, and platforms (e.g., energy consumption, and resource budget).

We research AI Agents to develop autonomous systems capable of complex reasoning, strategic planning, and active problem-solving through the integration of external tools. Our work focuses on orchestrating robust Multi-Agent Systems (MAS) where specialized agents collaborate to achieve high-level goals with precision and adaptability. A key area of our investigation is improving agent performance while strictly managing resource constraints; this includes the strategic transition from generalist large Language Models (LLMs) to efficient Small Language Models (SLMs) and the optimization of computational frameworks to maintain high performance. Furthermore, we are advancing interoperability through cutting-edge standards like the Model Context Protocol (MCP) for seamless tool integration and Agent-to-Agent (A2A) communication frameworks, enabling diverse agents to coordinate, negotiate, and execute tasks within a unified, self-evolving ecosystem. By treating these agents as "digital employees" capable of autonomous workflow planning and self-correction, we aim to deploy robust solutions across critical industries, ranging from automated compliance and fraud detection in FinTech to optimized predictive maintenance in manufacturing and hyper-personalized, 24/7 support in customer service.

Being able to accurately predict human mobility can inspire a lot of potential and promising applications, including location-based service networks, city planning, and infectious disease control. Therefore, the purpose of our work is to design low-cost prediction models that estimate the next locations and future encounters of the individuals with high accuracy. In order to achieve our goal, we analyze spatio-temporal characteristics along with social relationships and use them to construct prediction models based on machine learning techniques (such as support vector machines, feed-forward neural networks, and recurrent neural networks). 

We study mobile sensor network (MSN) architectures and distributed algorithms to monitor the moving phenomena in the disaster area in an open and unknown environment using a group of autonomous mobile sensor (MS) nodes. Monitoring a moving phenomenon is challenging due to the limited communication/sensing ranges of MS nodes, the phenomenon's unpredictable changes in distribution and position, and the lack of information on the sensing area. We propose distributed and robust algorithms in order to address these challenges and maximize the weighted sensing coverage.