Adaptive Broadening of Attention Guidance

Determine when and how much to broaden the human-attention guidance from underspecified features to well-optimized features during successive preference queries, and establish whether changing the attention instruction mid-task erodes ease of answering and understanding of the queries.

Background

The proposed attention mechanism directs the human exclusively toward features diagnosed as underspecified, so preference responses primarily inform those dimensions. The authors note that, as these features become resolved, later queries should also incorporate well-optimized features to help the robot learn how to balance objectives rather than focus on a single dimension.

The unresolved issue is how to schedule and control this expansion of attention, including the extent of broadening at each stage. The authors also identify a human-factors question: changing the instruction during a task may affect how easy the queries are to answer and how well participants understand their purpose. They state that resolving these questions requires substantial human validation.