Abstract:

Machine learning is a powerful paradigm which enables autonomous decision making by learning from examples. Despite its successes, human learning and decision making still vastly outperform autonomous decision making, particularly for complex sequential decision making tasks, where decisions made now have great ramifications far into the future. In this talk, I will present machine learning techniques with formal performance guarantees that efficiently learn to perform well in the apprenticeship learning setting---the setting when expert demonstrations of the (sequential decision making) task are available. I will also describe how my apprenticeship learning techniques have enabled us to solve real-world problems that could not be solved before. For example, they have enabled a helicopter to perform by far the most challenging aerobatic maneuvers performed by any autonomous helicopter to date. They have also enabled us to learn an autonomous controller for a quadruped robot to traverse challenging terrains and to learn a variety of different driving behaviours in our highway driving simulator.