Controlling Search Agents to Perform Search with Noisy Observations and Probabilistic Guarantees
P Thaker (Mitsubishi Electric Research Laboratories)
The core idea is a two-level bandit: one level maintains confidence bounds on per-region classifications, the other selects which feasible path to fly next by aggregating those bounds. What makes it usable in practice is that path selection respects fuel limits, start/end depots, and no-fly zones while still giving finite-time, probabilistic guarantees on when every region is classified.
Abstract
A control system and a method for controlling search agents to perform search with noisy observations and probabilistic guarantees is provided. The control system collects confidence bounds of a probabilistic classification of at least one region within at least one path of a set of paths. The control system compares aggregations of the confidence bounds of the probabilistic classifications of each path of the set of paths based on the collected confidence bounds, a first path of a set of paths is selected, for visit by a first search agent based on the comparison. The control system commands the first search agent to visit the selected first path to collect measurements associated with each region within the selected first path. The control system updates the confidence bounds of the probabilistic classifications of each region within the selected first path based on the measurements associated with the corresponding regions.
Read the explainer — intuition, the key idea, and honest limitations