Relation to Existing Work

Although this research will be based in part on Russell and Wefald's metareasoning techniques, it will go beyond that work in several ways.

• This work will focus on decision trees rather than game trees, leading to somewhat different details for the propagation of utilities.
• I intend to reevaluate some of the simplifying assumptions Russell and Wefald make, and make different and hopefully more general assumptions of my own.
• I plan to generalize their representations of probability distributions over future utilities.
• I plan to make an abstract characterization of the metareasoning problem, with a clear interface between the metareasoning strategy and the non-metareasoning parts of the algorithm, so that different metareasoning strategies can be plugged in and their effectiveness compared. I will investigate several alternative metareasoning strategies and compare them.
• I will focus on demonstrating that metareasoning is actually helpful. In particular, if the metareasoning overhead causes the system to perform poorly in comparison with a non-metareasoning evaluation strategy, that will be considered unacceptable. I hope to make a metareasoning module that can adjust the amount of time spent in metareasoning so as not to decrease overall performance.
• I hope to make a clean, elegant implementation of the system that will be useful for pedantic purposes.
• I hope to apply these techniques in a large, real-world domain such as medical decision-making as well as in a small, well-defined domain such as game-playing.

I hope to also make use of decision-theoretic control techniques developed by other researchers such as Pearl, Horvitz, Heckerman, Hansson and Mayer, and others. From preliminary investigations, much of that work seems very relevant.