An Embodied Conversational Agent (ECA) for Command & Control

In 2000/01 I was responsible for the dialogue manager for a large project to provide a virtual assistant in military decision making. At the time Microsoft's Paper Clip was all the vogue - sure it wasn't popular - but was it effective? Could it be adapted to a new domain where it might work better? This project introduced me to the ECA community who, unlike the NLP community, were actively interested in how humans interact with other humans and hence how we might make machines interact naturally with people. Their focus was on body language and emotion, but people such as ourselves and the group at Bielefeld working on a museum guide had real data and for us there is a more fundamental issue. The first thing to get right is to have the system provide mixed initiative at the discourse level. The is the terminology used in Kopp et al "A Conversational Agent as Museum Guide - Design and Evaluation of a Real-World Application" (Yorick you'll need to read the article - you can't just grep for the string). Mixed initiative is something the IVR community think they already have but what is needed is a system that can recognise when its conversational partner changes his or her mind about what they are trying to do. What is needed is a system that can handle mixed initiative at the level of a participant's intent. People changing their minds may seem rare; nope there are examples in the highly task oriented Communicator data. And it may seem hard; nope there is an obvious solution when you use a BDI architecture. The BDI (Belief Desire and Intention) model is one solution - there are no doubt others. BDI is often percieved as being a formal calculus of intent, and it is often percieved as being simply Good Old Fassioned AI, but as we were using it in its Rao and Georgeff incarnation, its primarly feature is that BDI manages an agent's commitment to a plan. But this is an old story...