Conformant Planning via Classical Planners

Conformant planning is the problem of computing a sequence of actions that achieves a goal in presence of incomplete information about the initial state (Smith and Weld 1998). Recent research shows that conformant planning could be very useful in the construction of finite-state controllers (Bonet, Palacios, and Geffner 2009) and in contingent planning (Albore, Palacios, and Geffner 2009). One of the most difficult issues, that directly affects the performance and scalability of conformant planners, is the size of the initial belief state—which is often exponential in the number of object constants of the problem. We observed that in many problems drawn from the recent International Planning Competitions (IPC) and from the literature, the initial belief states of many large instances contain more than 2 states, creating challenges to existing conformant planners. Various techniques have been developed to deal with the potentially huge size of the belief state. Some planners employ a compact representation of belief states—e.g., CFF (Brafman and Hoffmann 2004), POND (Bryce, Kambhampati, and Smith 2006), CNF (To, Son, and Pontelli 2010). Other planners develop simplification techniques that can reduce the size of the initial belief state, sometimes by several orders of magnitude—as in CPA (Tran et al. 2009) and DNF (To, Pontelli, and Son 2009). Most of these planners search for solutions in the belief state space. An alternative approach has been proposed in (Castellini, Giunchiglia, and Tacchella 2001) and (Kurien, Nayak, and Smith 2002), where the conformant planning problem is viewed as a set of sub-problems, which are classical planning problems, and solutions are computed using a two-step approach. CPLAN, developed in (Castellini, Giunchiglia, and Tacchella 2001), starts by computing a solution for a subproblem, called a possible plan, using a SAT-planner. It then checks whether the possible plan is a solution of the original problem. If it is not, the possible plan is discarded, a new possible plan is generated, and the process continues. FRAG-PLAN, proposed in (Kurien, Nayak, and Smith 2002), follows a slightly different approach in computing plans. It begins with the computation of a possible plan and then attempts to extend it to a valid plan. During the ex-