• Calendar

September 2, Wednesday
12:00 – 13:30

Agents operating in the real world need to handle both uncertainty and resource constraints. Typical problems in this framework are rational observation control (ROC) and optimal allocation of computation tasks during reasoning and search (also known as meta-reasoning). In both tasks, a crucial issue is value of information, a quantity hard to compute in general, and thus usually estimated using severe assumptions, such as myopic look-ahead, independence of information sources and separability of reward function.

Recently, a dynamic programming approach, which bypasses the myopic assumption, was introduced. An efficient method, based on this approach, constructs an optimal observation plan for a chain-shaped dependency model with exact measurements and additive reward function.

In this work we consider several extensions for the above method, that allow inexact and multiple measurements, various reward functions and more general dependency models. We developed a unifying approach to a wide class of ROC problems. To this end we extended the concept of a conditional performance profile, and developed an efficient technique for compiling a composite system beyond the input monotonicity assumption. The resulting framework can be applied to many real-world domains, such as medical diagnostics, setup optimization, mobile robot navigation and game-tree search.