Will my Spoken Dialogue System be a Slow Learner?

  • Layla El Asri ,
  • Romain Laroche

Proceedings of the 14th Annual Meeting of the Special Interest Group on Discourse and Dialogue (SIGDIAL) |

This paper presents a practical methodology for the integration of reinforcement learning during the design of a Spoken Dialogue System (SDS). It proposes a method that enables SDS designers to know, in advance, the number of dialogues that their system will need in order to learn the value of each state-action couple. We ask the designer to provide a user model in a simple way. Then, we run simulations with this model and we compute confidence intervals for the mean of the expected return of the state-action couples.