Human-Agent Teaming with Learning-Capable Agents
Abstract
In this human-agent teaming experiment, participants took the role of a robotic combat vehicle gunner working with a simulated, learning-capable target identification agent to classify potential threats encountered while patrolling through a simulated environment. Participants completed four trials. The agent would classify persons as encountered and the participant reviewed the agent classification and submitted the report. When the agent could not complete a classification, it inferred needed information from the participant's input to later update its programming. Transparency of agent learning was manipulated in three trials. In the fourth trial, the participant directly "taught" the agent rather than allow it to infer the needed information. Within-subjects repeated-measures analysis showed learning transparency supported performance, workload, trust, and perceptions of the agent. However, in the human-directed learning condition, task performance, workload, performance satisfaction, trust in the agent, and perception of the agent's reliability, animacy, likeability, and intelligence all worsened compared to the learning from inferred information conditions. Overall, these findings indicate that transparency of agent learning is an important design feature to include to support effective human agent teaming, but human-directed learning should be fielded with care, as the deleterious results could jeopardize mission effectiveness.
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 12, 2024
- Accession Number
- AD1224556
Entities
People
- Giovanna Camacho
- Jaquelyn Shreck
- Julia L Wright
- Justin Lee
- Shan G. Lakhmani
Organizations
- Oak Ridge Associated Universities
- United States Army Research Laboratory