Within Task Preference Elicitation in Net Benefit Planning

In order that an agent can be an effective collaborator it is
important that the agent is able to adapt its behaviour for the
preferences of a particular user. User preference elicitation
has been considered as a process that happens prior to plan
execution and typically prior to the planning process. However, when entering an interaction with a new human user
it will not always be possible or desirable for an elicitation
episode to take place. Moreover, the cost of any elicitation
(e.g., annoyance) must be weighed against its benefit in distinguishing between alternative plans. We therefore pose the
problem of within task preference elicitation, which explicitly represents the agent’s knowledge about the user’s preference model and how the agent’s knowledge can develop
as the interaction progresses. Our approach parameterises a
utility model for a net benefit planning task with a set of (observable) user attributes. This set of user attributes are represented as unknown values in a partially observable planning
model and can be accessed through guarded sensing actions
(e.g., through asking a question when it becomes relevant),
allowing the planner to reason with the possible alternative
user utility models. In this work we define the within task
preference elicitation problem and present our framework for
solving these problems. We present results examining its use
in modified benchmark scenarios, including a new planning
domain based on a tour guide scenario.