In this first instance, a feasible problem formulation is to be investigated (eg: as an optimisation problem), with some investigation into choosing an appropriate performance measures along with a priori known measurement models. For example, the Fisher information, a measure of adversary detectability. The agent (sensor) control problem will be cast in the multi-armed bandit setting, with actions corresponding to allocating limited resources as well as affecting the sensors (such as moving positions, changing focus, polling, or replacement). Since acquiring information may be limited not only to the number of the heterogeneous sensors but also the cost of different types of sensors in the environment (particularly those based on human sources), the model will incorporate resource supply limitations. These endow actions with budgets, making the optimisation of resources difficult. This problem will be addressed by estimating payoffs and resource consumption based on various rational decision-making models (eg: minmax, maxmin, etc), using confidence bounds for both payoffs and resource consumption.
Defence Science and Technology Group