Making agents learn tasks and controlling their autonomy

Hyuckchul Jung

My talk will present two topics "task learning" and "adjustable autonomy". First, task learning is to compose new tasks out of already known tasks, and problems include learning the sequences, control structures, parameter relationships, conditionals, etc. While many machine learning techniques have been developed, task learning is still a challenging problem. I will present an approach in which agents learn a high level task by interacting/discussing with a human user, and show the PLOW (Procedural Learning On the Web) system developed by the IHMC TRIPS group as a fully implemented task learning system.

Second, adjustable autonomy is the ability to dynamically control the boundary of agents' autonomy. While it is hardly a new topic in agent systems, it has been generally studied in a limited sense (e.g., transfer of control and permission/obligation). I will present new formalisms for adjustable autonomy (describing its multi-dimensional nature) and introduce a system called Kaa (KAoS adjustable autonomy) that provides policy-based capabilities to control agents' autonomy.