AI PLANNINGS STRONG SUIT

Planning systems generate partially ordered sequences of actions (or p lans) that solve a goal. They start from a specification of the valid actions (also called operators), which includes both the conditions un der which an action applies (the preconditions) and the expected outco me of applying that action (the effects). The general problem is quite hard both because of the potentially enormous search space and the di fficulty in fully and accurately representing real-world problems. App roaches to planning include operator-based planning, hierarchical task -network planning, case-based planning, reactive planning, and many mo re. Early planning work focused largely on ''toy'' problems (for examp le, the blocks world). More recently, there has been a big push toward applying planning systems to real-world applications. While planning systems have not yet achieved the level of commercial success enjoyed by some other areas of artificial intelligence-neural nets, for exampl e-a number of successful applications of planning technology to real-w orld problems have recently emerged. This installment of ''Trends & Co ntroversies'' highlights five such applications. I have asked the deve lopers of these systems to describe the application domain and the pla nning technology used to solve the problems. These systems all use som e farm of hierarchical task-network planning (in some cases combined w ith other techniques). HTN planning provides a way of specifying, as p art of the operator definition, how to hierarchically expand actions i nto partially ordered sequences (task networks) of actions. This appro ach succeeds, in part, because it provides a natural way of limiting t he possibly very large search spaces. See Readings in Planning (Morgan Kaufmann, 1990) or Artificial Intelligence: A Modern approach (Prenti ce Hall, 1995) for more details on various planning techniques. In the first article, Stephen Smith, Dana Nau, and Thomas Throop describe th eir use of planning technology to build a system for declarer play in contract bridge. The system can beat the best commercially available p rogram and is currently being incorporated into a commercial product. Second, John Mark Agosta and David Wilkins describe how the SIPE-2 pla n ner helps evaluate the US Coast Guard's ability to respond to marine oil spills. This system which automates a problem that is currently d one by hand, is undergoing evaluation by the Coast Guard. Third, Austi n Tate describes a planning application, in use by the European Space Agency, for the project management of spacecraft assembly, integration and verification. Fourth, Steve Chien and his colleagues describe the ir use of a planning system to automate the operations of NASA's Deep Space Network communication antennas. This system is currently being i ntegrated into a new system that will become operational in 1997. Fina lly, Thomas Lee and David Wilkins described their use of SIPE-2 in pro ducing military air campaign plans. Their planner is part of a demonst ration system that is fully integrated with the other software modules currently used for solving parts of this problem.