This paper considers centralized and decentralized control problems for partially-observed discrete event systems where sensor readings are assumed to be costly for reasons of bandwidth, energy, or security. The supervisory controllers, or agents, dynamically request sensors readings as needed to observe the trajectories of the system and correctly implement the given feedback control law. Thus, each sensor may be turned on/off several times along a given system trajectory. Different policies for dynamic sensor activation can be used by the agents. A set of policies is said to be minimal if any strictly less activation prevents the correct implementation of the control law. A systematic formulation of the dynamic sensor activation problem is proposed and its solution developed when the solution space is restricted to the set of transitions of a given automaton model of the system. Two algorithms that compute minimal sensor activation policies are presented, one for ensuring observability and the other for ensuring coobservability; observability and coobservability are key properties that arise in the solution of supervisory control problems for centralized and decentralized discrete event systems. These algorithms are of polynomial complexity in both the number of states and the number of events of the system.