Session “Adaptivity”

• Participants: Casey, Catha, David, Hendrik, Jens, Klaus, Phillip, Pierre, Srini,
• Notes taken by Hendrik

What is adaptivity?

• Adaptivity to what?
• Context vs. user: There is a difference between context-aware systems and user-adaptive systems.
• Do we have to think of adaptive behaviour in terms of prediction?
• Task vs. adaptivity: A train ticket systems that sells the user the requested ticket is certainly not adaptive.
• David: An adaptive system optimises/changes parameters to a specific condition (i.e., there are a lot of parameters and only a few conditions).
• Alignment is not adaptation.

• Where is more variation? Between users (interpersonal variation) or between the same user in different situations (‘inter-situational) variation)?

Approaches

• Where does the adaptivity happen? In the user model? In the sensors?
• Adapting and learning during one interaction vs. adapting and learning over time (i.e., online learning vs. offline learning).
• Different approaches: treat all users the same vs. build user groups, adapt to user by using group specific stereotypes vs. adapt to the specific user.
• Adapting the model to the signal vs. adapting the signal to the model
• Small control/feedback loops (in the cybernetics sense) have been implemented on the production side, e.g., by Roger Moore's TTS that listens to itself

Difficulties

• What is to be optimised? The optimisation goal is far from obvious for complex, situated systems (such as service robots).
• It is difficult to adapt to a potentially huge number of variables at the same time - you don't know where the variation comes from.
• What is a good metric for adaptation? Optimising module for module independently doesn't guarantee that the resulting behaviour is optimal.

Real world SDS

• Should a system be predictable to the user or should the system be able to predict what the user is doing?
• People might want systems to be static and predictable (Jens’ example of the Stockholm traffic information system that he does not want to change).

How to evaluate an adaptive dialogue system?

• Task efficiency vs. user satisfaction as evaluation goal: depends on task and application. Example: a complaint answering dialogue system doesn't care about people hanging up (or at least about not being helpful); a system that wants to sell should care.

Who should adapt? The user or the system?

• Dialogue systems should not behave like ‘jerks’ (i.e., that make the user adapt to them).
• Users could, however, be gently pushed into adaptive behaviour by the system using subtle cues. Example: Lowering the speech rate to make the user speak slower, this improves speech recognition rate (has been done; reference anyone?).
• Systems could be capable of both ‘jerky’ as well as adaptive strategies and then choose during the interaction how to behave. Example: A native speaker of German uses a dialogue system that has optimised models for English but only mediocre models of German. If the user’s Enlish is better than the systems German, the system should make the user adapt (i.e., speak English). If the user, however does not speak Enlish at all, the system should adapt to the user.
• When should the system adapt to the user, when should it make the user adapt? This depends on the role of the system (e.g., foreign language tutoring SDS); social goals; engineering issues and system adaptation costs, e.g., how difficult would it be for the system to adapt. This needs to be flexible depending on what makes sense.