Abstract:
Robotic algorithms rarely come perfectly pre-configured, and when choosing parameters, tradeoffs must often be made: between performance and robustness; efficiency and safety; the comfort of the user and the comfort of bystanders. While engineers can tune parameters by hand or carefully design reward functions to optimize over, this is not always a straightforward task. When the behavior of a system isn’t easily quantified, when its utility is defined by user acceptance, or when non-roboticists must be able to personalize their own experiences, we turn to learning from people.

In this thesis, I discuss planning and safety for assistive guide robots, an example of an area where gathering user requirements—another form of learning from people—is critical but difficult and personalization is necessary. I demonstrate the efficacy of the popular CMA-ES optimization algorithm, which requires only ranking information for fast convergence, for parameter tuning from ordinal (preference) information. Based on the theoretical foundations of the CMA-ES, I discuss a multimodal algorithm that can find multiple satisficing parameter sets from preferences.

Thesis Committee Members:
Aaron Steinfeld (Chair)
Changliu Liu
Andrea Bajcsy
Laurel Riek (U.C. San Diego)