Ever wondered how often the engine of your car should really be checked on defects? Or how often an electicity provider should check their cables to minimize both the chance of outages and their maintenance costs? Or how often a drone should use it’s battery-draining GPS-system to keep an accurate idea of it’s positions?
All these questions can be formulated as active measure problems, or AM problems for short, in which agents need to determine whether the value of taking a measurement is worth it’s cost. Problems that contain partial observability, as is the case here, are often modeled as partially observable Markov decision processes (POMDPs), but these are generally hard to solve for larger real-life problems. For this reason, we are interested in ways of moddeling AM problems in a way that makes solving them easier.
ACNO-MDPs and Act-then-measure
Prior work has focussed on active measure environments where all measurements are complete and noiseless (missing reference). In such settings, the