Rhai

Rich Human-Agent Interaction

Problem

Just like a human assistant, an agent needs to consult its supervisor when asked to perform a task that is under-specified, has ambiguous instructions, deviates from normal, or has changed.

The Rich Human-Agent Interaction (Rhai, pronounced "ray") project is exploring how to build good interfaces for interaction between a user and an agent. How should an agent ask a user to check its understanding of natural language? How should an agent ask a user to approve, reject, or modify actions that it proposes?

We are using calendar scheduling to answer these questions.

Solution

The RhaiCAL (Rich Human-Agent Interaction for Calendaring) system provides novel visualizations and interaction techniques for interacting with an intelligent agent, with an emphasis on calendar scheduling.

After an agent interprets natural language containing meeting information, a user can easily correct mistakes using RhaiCAL's clarification dialogs, which provide the agent with feedback to improve its performance.

When an agent proposes to take actions on the user's behalf, it can ask the user to review them. RhaiCAL uses novel visualizations to present the proposal to the user and to allow the user to modify or reject the proposal. The agent is informed of the user's actions in a manner that supports long-term learning of the user's preferences.

We have designed a high-level XML-based language that allows an agent to express its questions and proposed actions without mentioning user interface details, and that enables RhaiCAL to generate high-quality user interfaces.

Radar

The Rhai project is part of RADAR, which aims to build intelligent personal assistants that help people to efficiently handle common office tasks such as managing their calendars, handling email, writing reports, maintaining web sites, allocating project space, and so forth.

Publications

I presented an Interactive Poster at CHI 2005 describing this work:

Andrew Faulring, and Brad A. Myers. (2005). Enabling Rich Human-Agent Interaction for a Calendar Scheduling Agent. In Proceedings of the Conference on Human Factors in Computing Systems Extended Abstracts (CHI 2005). Portland, Oregon, USA, April 2–7, 2005. [Local PDF]