In supervisory control, the objective of observation is to guarantee a correct control decision. To observe an event occurrence, an associated sensing device must be activated, which incurs a cost. In this paper, an online algorithm is developed to minimize sensor activation while ensuring that the collected information is sufficient. In previous work, it was determined that observation problems can be reduced to distinguishing certain pairs of states. Now, this is extended by taking into account future system evolution. Applying the extended result, the online algorithm only needs to look one step ahead and remember the current state estimate. (C) 2016 Published by Elsevier Ltd.