<?xml version="1.0" standalone="yes"?>
<Paper uid="W00-0304">
  <Title>NJFun: A Reinforcement Learning Spoken Dialogue System</Title>
  <Section position="6" start_page="19" end_page="19" type="concl">
    <SectionTitle>
5 Limitations
</SectionTitle>
    <Paragraph position="0"> The main limitation of this effort to automate the design of a good dialogue strategy is that our current framework has nothing to say about how to choose the reward measure, or how to best represent dialogue state. In NJFun we carefully but manually designed the state space of the dialogue. In the future, we hope to develop a learning methodology to automate the choice of state space for dialogue systems.</Paragraph>
    <Paragraph position="1"> With respect to the reward function, our empirical evaluation investigated the impact of using a number of reward measures (e.g., user feedback such as U4 in Figure 1, task completion rate, ASR accuracy), and found that some rewards worked better than others.</Paragraph>
    <Paragraph position="2"> We would like to better understand these differences among the reward measures, investigate the use of a learned reward function, and explore the use of non-terminal rewards.</Paragraph>
  </Section>
class="xml-element"></Paper>