UC Berkeley

Human beings are decision making engines. From low level automatic activities, such as walking, to high-level reasoning, we readily solve complex perception and decision problems in real time. On the other end of the spectrum, computers and robots have proven particularly useful in dull, dirty, and dangerous domains---inspecting nuclear disaster sites, solving vast optimization problems, and generally accomplishing tasks difficult for humans. Recent advances in robotics and computer science have only served to underscore the difference between humans and robots, namely that robots tend to excel at lower level structured tasks, while humans are unmatched at higher level decision making, particularly when problems involve unstructured data, soft constraints, and ambiguous objectives.

This thesis explores decision problems that lie in the gap between current robot capability and human ability. A generic framework for representing large-scale decision problems, referred to as the Generalized Planning Problem (GPP), is defined. A learning-based GPP solution method is presented that captures multiple stochastic problem elements and soft constraints, which arise in many real world problems. An example routing problem is presented to showcase the expressiveness of the GPP framework. To leverage human intuition with such problems, the GPP framework allows for incorporating human input in a structured way as an external reward signal. Adding human feedback into the learning process results in a variably autonomous decision making engine that can vary continuously between fully autonomous, fully manual, and everywhere in between.