Computer Science and Engineering Division
Many
human-computer interaction (HCI) systems are sequential interaction systems in
which the designer has incomplete and uncertain knowledge about the system's
environment and in which user-feedback is impoverished, noisy, and delayed in
time. These are precisely the sort of problems reinforcement learning (RL)
methods are good at solving. In this talk, I will discuss the opportunities and
challenges facing the use of RL as a rigorous design principle for HCI, and
illustrate my arguments using examples from 3 simple RL-based HCI systems that
I have helped build: an adaptive spoken-dialogue system, an interactive
software agent in an online community, and most recently an adaptive reminder
system in a cognitive orthotic domain.