Towards Undifferentiated Cognitive Agents: Determining Gaps in Comprehension

Description: 

PI: Pascal Hitzler

This is a collaborative effort with the Air Force Research Laboratory (PIs Christopher Myers, Adam Stevens, Benji Maruyama) and Drexel University (PI Dario Salvucci).

Autonomous systems are a new frontier for pushing sociotechnical advancement. Such systems will eventually become pervasive, involved in everything from manufacturing, healthcare, defense, and even research itself. However, proliferation is stifled by the high development costs and the resulting inflexibility of the produced systems. The current time needed to create and integrate state of the art autonomous systems that operate as team members in complex situations is a 3-15 year development period, often requiring humans to adapt to limitations in the resulting systems. A new research thrust in interactive task learning has begun, calling for natural human-autonomy interaction to facilitate system flexibility and minimize users’ complexity in providing autonomous systems with new tasks. In this project, we respond to Laird, et al.’s (2017) call through the development of a cognitive system capable of independently acquiring most required task knowledge and skill to perform a set of tasks.

Our proposed approach begins with the development of a foundational and generalizable cognitive system that can be transformed into a specialized cognitive agent through written instruction, interactions with its trainers, task experience, and developer intervention (when needed). Specifically, we seek to research and develop an undifferentiated agent (uAgent) that is a set of general-purpose computational cognitive capacities enabling it to read task instructions, iteratively interact with trainers to fill gaps in task knowledge, generate the requisite task knowledge from the instructions, and complete multitasking scenarios of varying complexity. We propose to further specialize the agent to desired levels of task proficiency using the Autonomous Research System (ARES).

This material is based upon work supported by the Air Force Office of Scientific Research under award number FA9550-18-1-0386

Funding Agency: 

AFOSR

From: 

July, 2018

To: 

July, 2021