The LISTEN Reading Tutor can be thought of as an agent acting
under uncertainty. Its estimate of the current "state," that is the current
reading ability of the child, depends on noisy probes or observations such as
answers to automated questions presented to the child and measurements of latency
and production time as the child tries to read a word. Further, its interventions
or actions such as word building exercises, vocabulary help, etc., have uncertain
results. In the face of this uncertainty, the reading tutor needs to make
decisions about which interventions to use at each point in time in order to
maximize the child's progress in learning to read.
This project aims at addressing the second aspect, namely,
evaluating the effects of tutorial interventions on a lexical level.
The goal is to determine the long term expected reward of executing a given tutorial
intervention in a given tutor state, using reinforcement learning.