Research
Current work concentrates on two axes. Neither adds a layer to an existing method; both start by getting the question right.
Axis 1: AI Agents & Educational Innovation
Large language models have made tool-using, self-decomposing agents practical as teaching instruments, but placing an agent in a classroom opens two gaps at once. Learning can be hollowed out — a student outsources judgement to the model and submits work that looks complete but taught them nothing. And the work becomes unauditable: an agent leaves no trace of its intermediate decisions, so an instructor cannot tell which judgements were the student's own.
The question this axis addresses is when a workflow should be fixed in advance, buying reproducibility and accountability, and when decomposition should be handed back to the model, buying generalisation across unfamiliar task shapes. These are not successive generations of one idea; they purchase different properties, and the choice has to be argued each time.
The current approach builds controlled AI scaffolding into courses that are actually taught, records the interaction traces, and compares learning outcomes across scaffolding strengths using a quasi-experimental design. This is an active direction with no published results yet.
Axis 2: Intelligent IoT & Resilient Systems
The availability of a communication or sensing system is not captured by its average-case performance. What decides whether the system is usable is how much function survives degradation — when base stations are damaged, nodes run out of energy, traffic surges, or sensing is disturbed by the environment.
This axis breaks resilience into measurable problems: traffic offloading driven by user mobility in disaster scenarios; where to deploy chargers in a wireless rechargeable sensor network so coverage survives an energy budget; detecting anomalous flows in software-defined networking before an attack becomes widespread; and predicting environmental variables such as visibility, which bear directly on operational safety, from observational data.
The shared structure is to define the degraded scenario first, ask how much function is retained under it, and only then optimise.
Related publications
- H.-M. Chuang, C.-H. Tsai, and W.-K. Soong, LSTM-based hybrid deep learning models for visibility prediction: a data-driven approach, Advances in Meteorology, 2026. doi:10.1155/adme/3091260
- H.-M. Chuang, F. Liu, and C.-H. Tsai, Early detection of abnormal attacks in software-defined networking using machine learning approaches, Symmetry, 2022. doi:10.3390/sym14061178
- A.-H. Tsai and C.-H. Tsai, The hybrid traffic offloading mode for disaster-resilient communication networks based on user mobility, Wireless Communications and Mobile Computing, 2021. doi:10.1155/2021/9403982
- W.-P. Chen, A.-H. Tsai, and C.-H. Tsai, Smart traffic offloading with mobile edge computing for disaster-resilient communication networks, Journal of Network and Systems Management, 2019. doi:10.1007/s10922-018-9474-z
- K.-H. Yao, J.-R. Jiang, C.-H. Tsai, and Z.-S. Wu, Evolutionary beamforming optimization for radio frequency charging in wireless rechargeable sensor networks, Sensors, 2017. doi:10.3390/s17081918
Earlier trajectory
These lines explain where the current directions came from. They are not the present focus.
Mobile augmented reality and context awareness (2006–2018)
Earlier work centred on letting a system infer a user's current situation — improving GPS positioning with context-aware techniques, switching augmented-reality content according to user behaviour, and using mobile AR as a scaffolding platform for outdoor fieldtrip learning. Published in Communications of the ACM, Computer Standards & Interfaces, and Journal of Internet Technology.
Image matching and video foreground detection (2015–2018)
A second line addressed measurement problems in imagery: accelerated matching and registration of UAV orthoimages, real-time foreground detection under changing environments via adaptive ensemble learning, and detecting abnormal crowd behaviour from foreground features. Published in the ISPRS Journal of Photogrammetry and Remote Sensing, Information Fusion, and the Journal of Electronic Imaging.
Simulation environments and immersive interaction (2006–2025)
Work in this lineage covered decision-support architectures for distributed simulation environments, networked mixed-reality campuses, and more recently the key technologies and involvement measurement of metaverse performance systems. It is not a primary focus at present.