Adaptive machine learning for changing environments

Explaining the science

Efficient reinforcement learning

Unlike supervised or unsupervised learning based on large amounts of unified and stationary datasets, reinforcement learning focuses on how an agent should continuously interact with its environment to maximise its reward. Under such a setting, massive trials are generally necessary before or during the learning process, and success still highly depends on manually crafted learning architectures and targets. This project aims at solving the sample inefficiency problem in existing approaches.

On-line learning

'Off-line' machine learning approaches typically learn from well-organised and stationary large-scale datasets through batch processing. In stream-oriented, sensor-based systems that are continuously delivering data, this would be unrealistic and inappropriate for many real-world infrastructure applications. Ideally, the learning of targeted hidden knowledge should be incremental, leveraging multi-source data streams. This project aims at developing online algorithms for continuous learning with guarantees of success.

Small-sample-size adaptive learning

Existing machine learning approaches generally train their models with massive amounts of data collected in stationary environments. However, such approaches are not applicable to the ever-changing environments typically found when sensor/actuator systems are deployed to make an infrastructure smart (e.g. smart cities, autonomous vehicles etc.). It's unfeasible to re-collect and re-organise the data and re-train the previously learnt models whenever the environment changes. This project aims at developing high level 'meta-learning' algorithms that can rapidly notice environmental changes based on a limited amount of sensing data samples, and continuously adjust the rest of the learning model accordingly.