Abstract Milica Gašić May18 2018
In spoken dialogue systems, we aim to deploy artificial intelligence to build automated dialogue agents that can converse with humans. As part of this effort, we need to find ways to optimise the dialogue policy, i.e. we need to optimise a function that takes the current state of the dialogue as input and returns the response of the system. This is normally done via reinforcement learning. Deep reinforcement learning approaches have produced state-of-the-art results on games. In this talk I will discuss the necessary steps needed to deploy deep reinforcement learning for dialogue policy optimisation. I will also discuss the necessity for common benchmarks and the efforts in the Dialogue Systems Group to provide these.