Barbara Grosz Concludes 2008 Big Thinkers Series

The final talk of the Yahoo! Research 2008 Big Thinkers Series featured Barbara Grosz, Higgins Professor of Natural Sciences at Harvard University. Entitled “It's Time to Talk: Managing Interruptions in Multi-Agent Environments,” her talk focused on looking at computer systems as collaborative partners or teammates, rather than simple servants that just respond to queries, and how a system can prompt a human for information with the consequence of interrupting their workflow when it believes that the human has extra information beyond the system.

Grosz began by speaking about the concept of teamwork. “When I started, the protocol was a single person using a single computer,” said Grosz. “Now, they are predominantly used in networks: many people with many systems working together on the Web to get a job done.” She illustrated this concept through an example of a rescue team that is constantly communicating and collaborating.

Grosz then introduced the notion of interruption. Several shots of computer dialogue boxes appeared on her presentation, with messages like “Fatal error: unexpected catastrophic failure – investigate” or “Cancel your reservation?” with 2 choices presented: “Cancel” or “Ok” to which Grosz remarked that the choices were confusing. Understanding when to interrupt is one kind of decision that a system has to make among a whole host of decisions in a collaborative system.

An example Grosz illustrates is WAID – a collaborative interface that helps people with citations while writing papers. Typically, writers of papers are interrupted when citing sources. With WAID, a search engine looks for citations while the writer continues to write without interruption. The writer gives search terms to the system while writing and the system searches the Web, libraries and other sources and returns a set of solutions. This exemplifies a collaborative system that considers which team member is good at what – the labor is then divided accordingly. “Collaborative interfaces exist where the system and person work together to solve a problem that neither could solve alone,” said Grosz. Although the system never interrupts, Grosz raised the question of whether or not there are times when it calls for interruption. Computers continually make decisions on when to interrupt - needing to make these decisions and having the right information available is important.

To further research multi-agent decision making, Grosz and colleagues at Harvard and Bar Ilan University developed a testbed game called Colored Trails (CT). It consists simply of a grid of different colored squares with one or more squares designated as the goal square for each player. Each player has a piece on the board in one of the non-goal squares and a set of colored chips. A piece can be moved into an adjacent square only if the player has a chip of the same color as the adjacent square. The player’s objective is to get to the goal. The game framework allows researchers to specify different scoring rules, and thus to vary the extent to which individual and social good are rewarded.

Grosz and her group at Harvard defined a two-player (a human and a computer) CT-interruption game in which goals moved stochastically. Only one player, the human player, received information about the new locations of the goals as they moved. The computer player could only obtain this information by interrupting the person. If an interruption is accepted, the players lose one turn moving their pieces. Although the second player was always a computer, the experimental conditions varied whether subjects were told they were playing a person or a computer.

The experiment Grosz reported on yielded two conclusions: players are more likely to accept interruptions if the outcome is higher; and the perceived partner type (human vs. computer) affects people’s responses only some of the time. As the value of the interruption goes up, so does the likelihood of acceptance. If the value is either low or high, the partner type (human vs computer) did not affect the willingness of a person to accept an interruption. For intermediate level values, people were more willing to accept interruptions from a (perceived) person rather than a computer.