Position: PhD Candidate
Current Institution: Massachusetts Institute of Technology
Abstract: Towards Integrated Intelligent Systems
Throughout the history of artificial intelligence (AI), it has been a common theme to fragment the grand goal of building intelligent agents into many in-depth but often incompatible subproblems, including statistical machine learning, optimization, motion planning, computer vision, etc. My research stems from a different perspective that emphasizes the breadth: how to integrate those individual modules into a cohesive intelligent system that can act and adapt itself in complex environments.
My work focuses on one of the most challenging aspects of the integration of AI modules: learning and planning. Learning aims to represent knowledge as a model, while a planner makes decisions. This problem is a variant of the reinforcement learning problem, but the practicality of classic methods is doubtful for long-horizon high-dimension problems. A central question of integrating learning and planning is how to decide which actions to select from a potentially infinite action space. The choice of actions determines both the performance of the planner and the data to be gathered for learning.
To this end, I have worked on the mathematical formulation of action selection as Bayesian optimization, active model learning for structured actions, and task and motion planning that enables long-horizon planning with learned models. Currently, I am investigating how to integrate perception, learning, and planning, and control modules to demonstrate intelligent behaviors on a physical robotic system. In particular, I would like to understand how to “wire” the existing modules together so that desired properties such as tractability and analyzability can be induced.
Zi Wang is a PhD candidate at the MIT Computer Science and Artificial Intelligence Laboratory, advised by Stefanie Jegelka, Leslie Kaelbling, and Tomas Lozano-Perez. Her research interests lie broadly in machine learning and artificial intelligence, currently with applications to robotics problems.
Zi received a master’s degree in electrical engineering and computer science from MIT in 2016, and a bachelor’s degree in computer science and technology from Tsinghua University in 2014. She has received many awards, including the Greater China Computer Science Fellowship and the Google Anita Borg Scholarship. while at MIT, she served as a co-president of Graduate Women in Course 6 (EECS) and as a reviewer of many leading conferences and journals in her field, such as the Annual Conference on Neural Information Processing Systems (NIPS) and the Journal of Machine Learning Research (JMLR).