In this talk I will provide an overview of the rapidly growing field
of Statistical Learning for adaptive Multimodal Dialogue Systems. I
first provide some background on Dialogue Systems, their architecture,
and (some of) the current research challenges in the field.  I argue
for the use of statistical methods for dialogue control, as they lead
to more robust and adaptive interaction.

I focus on two application examples. First, I introduce a Supervised
Learning approach to re-ranking dialogue prompts based on predicted
Text-To-Speech quality. Then I introduce a 5-step method to
"bootstrap" Reinforcement Learning-based dialogue behaviour. We apply
this method to optimise Dialogue Management and Natural Language
Generation. We evaluate all the results with real users.