Zizhao Wang

I am a PhD student in ECE at the University of Texas at Austin, advised by Prof. Peter Stone. My research focuses on the intersection of causality and decision-making (mostly reinforcement learning).

Previously, I completed my M.S. in CS at Columbia University, advised by Prof. Peter Allen and Prof. Itsik Pe’er, and my undergraduate studies at the University of Michigan - Ann Arbor.

Email  /  Scholar  /  Twitter  /  CV

Selected Works

A unsupervised skill discovery method that induces diverse interactions between state factors, which are often more valuable for solving downstream tasks.

Improve the robustness of motion prediction models by inferring causal relationships between agents.

A method for learning disentangled skills where each skill component only affects one factor of the state space, so skills can be efficiently reused to solve downstream tasks.

Improve exploration by visiting states where the agent is uncertain whether (as opposed to how) entities such as the agent or objects have some influence on each other.

Learn the causal relationships in the dynamics and reward functions for each task to derive a minimal, task-specific abstraction.

When task and motion planning search reaches dead-ends due to an early action, we identify and backjump to the culprit action to resample it, improving the search efficiency compared to backtracking.

Learn a theoretically proved causal dynamics model that removes unnecessary dependencies between state variables and the action, so it generalizes well to unseen states.

Generate cheap training data for navigation by hallucinating obstacles.

Learn how to dynamically adjust planner parameters using evaluative feedback from non-expert users.

Learn how to dynamically adjust planner parameters using interventions from non-expert users.

A novel variational inference framework for nonlinear hidden Markov models.

Teaching and Services