The design of MOO agents: Implications from a study on multi-modal collaborative problem solving.

2. Empirical study: collaboration in a multi-modal virtual environment.

We report experiments on collaborative problem solving in a virtual environment. Our long term goal is to develop computational models for supporting human-computer collaboration (Dillenbourg, 1996). The specific goal was to study how two human subjects elaborate a shared solution of a problem and how the drawings they make on a shared whiteboard help to support mutual understanding.

In the experiments, the two subjects play detectives in a mystery solving game: Mona-Lisa has been killed and they have to find the killer. They walk in a text-based virtual world (a MOO environment) where they meet suspects, ask questions about relations with the victim, regarding what they have done the night of the murder, and so forth. Suspects are programmed robots. When exploring rooms, they find various objects which help them to discover the murderer. They are told that they have to find the single suspect who (1) as a motive to kill, (2) had access to the murder weapon and (3) had the opportunity to kill the victim when she was alone.

The subjects are located in different places, connected by a computer network and communicate using two pieces of software: a text-based MOO system and a Whiteboard. The MOO environment is a standard MOO called TECFAMOO[2]. The subjects move in different rooms, talk to each other via two commands: "say...", to communicate with anybody in the same room, and "page John..." to communicate with John where ever he is. The detectives each carry a notebook which automatically records the answer to all the questions that they ask. In this experiment, the subjects use a MOO client called TKMOOlight which runs on UNIX stations. It includes an elementary whiteboard: both users draw on a same page, can see and edit the objects drawn by their partner, but they do not see each other's cursor. All actions and interactions in the MOO window and in the whiteboard are recorded.

We ran the experiments with 20 pairs, the average time was two hours. A more complete analysis of observations can be found in Dillenbourg and Traum (1997). We present here only a few particular results which seem interesting for the design of artificial agents. In the following presentation, we refer to four levels of mutuality of knowledge (from Clark, 1994): (1) A knows that B can access knowledge X; (2) A knows that B has perceived X; (3) A knows that B understood X more or less as A understood it, and (4) A knows that B agrees on X. These four indented levels cover both the grounding process (achieving mutual understanding) and the negotiation process (acceptance or agreement.


[2] Visit tecfamoo.unige.ch - port 7777 For information: http://tecfa.unige.ch/tecfamoo.unige.ch

The design of MOO agents: Implications from a study on multi-modal collaborative problem solving - 21 MARCH 1997

Generated with Harlequin WebMaker