Sriraam Natarajan, Kshitij Judah, Prasad Tadepalli, Alan Fern
There is a growing interest in intelligent assistants for a variety of applications from organizing tasks for knowledge workers to helping people with dementia. In our earlier work, we presented a decision-theoretic framework that captures the general notion of assistance. The objective was to observe a goal-directed agent and to select assistive actions in order to minimize the overall cost. We employed the use of POMDPs to model the assistant whose hidden state was the goal of the agent. In this work, we evaluate our model of assistance on a real world domain and establish that our model was very effective in reducing the efforts of the user. We compared the results of our model against a cost-sensitive supervised learning algorithm. We also describe our current work on extending the model to include relational hierarchies. We then analyze some problems in our model and suggest possible extensions to handle them.
Subjects: 15.5 Decision Theory
Submitted: Jan 26, 2007