Leap Before You Look: Information Gathering in the PUCCINI Planner

Most of the work in planning with incomplete informa tion takes a look before you leap perspective Ac tions must be guaranteed to have their intended e ects before they can be executed We argue that this ap proach is impossible to follow in many real world do mains The agent may not have enough information to ensure that an action will have a given e ect in advance of executing it This paper describes puccini a partial order planner used to control the Internet Softbot Et zioni Weld puccini takes a di erent approach to coping with incomplete information Leap before you look puccini doesn t require actions to be known to have the desired e ects before execution However it still maintains soundness by requiring the e ects to be veri ed eventually We discuss how this is achieved using a simple generalization of causal links