IQ-adaptive statistical dialogue management using Gaussian processes

Adapting a Spoken Dialogue System to the user's satisfaction is supposed to result in more successful dialogues. In this thesis, Gaussian processes are used to model a policy for a statistical Spoken Dialogue System and the Interaction Quality (IQ) metric which is a measure for the user's satisfaction is used to train this policy. As the policy decides which actions are taken next at a particular point, the dialogue flow is thus adapted to the IQ. Afterwards, it is investigated whether the incorporation of the IQ metric is beneficial. Therefore, different learning strategies with and without the IQ metric are used to train different policies. Then, the performance of all trained policies is evaluated regarding dialogue completion, task success, the average length of a dialogue and the average IQ value at the end of a dialogue