Dr Ming Hay Chung

PhD thesis:

Learning to Plan with Reinforcement Learning

My PhD project focuses on how to train AI systems to learn to plan through interaction with the environment, without relying on handcrafted planning algorithms. Despite the success and prevalence of handcrafted planning methods such as AlphaGo, these approaches are largely limited to restricted domains, such as board games. Learning to plan aims to make planning more general, closer to the flexible and learnable way in which humans and animals plan. I believe that learning to plan is also essential for general intelligence, as handcrafted planning methods often fail to adapt to the complexity of real-world environments.

The project led to novel algorithms that achieved new state-of-the-art performance in AI planning domains, and the results have also been extended to large language models. This work resulted in several publications at top machine learning conferences and has inspired further research studying and improving planning in modern AI systems.

After the PhD

I started a start-up called Dualverse AI (dualverse.ai) after completing my PhD. Since my doctoral research also focused on building better environments for AI systems, I extended this direction to a multi-agent setting that enables AI agents to pursue scientific discovery.

We proposed the Station: a miniature research world in which AI agents can discuss ideas, run experiments, publish papers, and follow their own research journeys, imitating aspects of how human scientific communities work. The Station has shown promise in driving scientific discovery, including advancing the frontier on several benchmarks and solving open mathematical problems that other AI systems have not solved.