Sensor scheduling for hunting elusive hiding targets: a restless bandit index policy

We consider a sensor scheduling model where a set of identical sensors are used to hunt a larger set of heterogeneous targets, each of which is located at a corresponding site. Target states change randomly over discrete time slots between “exposed” and ‘hidden,” according to Markovian transition probabilities that depend on whether sites are searched or not, so as to make the targets elusive. Sensors are imperfect, failing to detect an exposed target when searching its site with a positive misdetection probability. We formulate as a partially observable Markov decision process the problem of scheduling the sensors to search the sites so as to maximize the expected total discounted value of rewards earned (when targets are hunted) minus search costs incurred. Given the intractability of finding an optimal policy, we introduce a tractable heuristic search policy of priority index type based on the Whittle’s index for restless bandits. Preliminary computational results are reported showing that such a policy is nearly optimal and can substantially outperform the myopic policy and other simple heuristics.