Automated planning technology has become mature enough to be useful in applications that range from game playing to control of space vehicles. In this article, Dana Nau discusses where automated-planning research has been, where it is likely to go, where he thinks it should go, and some major challenges in getting there. The article is an updated version of Nau's invited talk at AAAI-05 in Pittsburgh, Pennsylvania. One motivation for automated-planning research is theoretical: planning is an important component of rational behavior--so if one objective of artificial intelligence is to grasp the computational aspects of intelligence, then certainly planning plays a critical role. Another motivation is very practical: plans are needed in many different fields of human endeavor, and in some cases it is desirable to create these plans automatically.

First, besides dealing with time constraints and other constraints of the job shop domain (for example, equipment constraints), crew planners must also deal with space constraints to prevent space discontinuities in duties, positioning crew where they are needed, whether as passengers in trains or other transportation means. Second, crew planners must also deal with complex train frequencies, such as week frequencies (for example, a train might only run on weekends), year periods (for example, only during summer), and special days (for example, on holidays and days before holidays), which put additional constraints on the combinations of tasks. These two aspects are critical to the quality of the final schedules and the efficiency of the scheduling process, requiring abstraction techniques to be used extensively. Third, the work periods of crew do not have fixed times, as shifts in industry, but can slide during the day to accommodate the irregularity of the train operation, although subject to constraints. The resource-sliding dynamics make it difficult to analyze activity demand and resource contention as is usually done in the job shop domain (Sadeh and Fox 1991).

With the exception of plan repair, important topics related to the use of plans (robust execution, reactivity, monitoring, evaluation) have received significantly less consideration. In realistic domains, however, plan generation is only a small component of the overall package. In particular, plans must be updated in response to new information and requirements in a timely fashion to ensure that they remain viable and relevant. Plan execution involves more than blind adherence to previously generated plans. Rather, run-time decisions are made to adapt, initiate, or abandon plans and activities in response to current considerations within the operating environment.

DSIPE supports a human planner. Although their requirements overlap for this mission, they each have independent goals (other missions to be performed or supported), capabilities, and resources. We extended SIPE-2's internal plan representation Throughout the planning process, each planning system monitors the local cell's planning activity for constraints and subgoals that might be relevant to other planning cells and notifies the cells of this information. For example, the naval planning cell might notify the Marine Corps planner that a particular landing area will be swept of mines by a specified time. Currently, the only constraints that are monitored in this way are the postconditions.

Introduction One of the major deterrents to productivity in industry today is the inability to effectively manage and control production. The problem is particularly acute in job shop environments where plant operation is routinely characterized by high work-in-process (WIP) inventories, tardy orders, poor resource utilization, and other shop floor inefficiencies. Perhaps the single most significant obstacle to improved factory performance is the complexity associated with constructing and maintaining good production schedules. Good schedules must reflect both the full detail of the operating environment and the influence of a conflicting set of preferences that range from global organizational objectives to specific operational idiosyncrasies. Existing computer-based techniques for production scheduling are capable of incorporating only a small fraction of this scheduling knowledge and, as a result, typically produce schedules that bear little resemblance to the actual state of the ...

The Workshop on Comparative Analysis of AI Planning Systems, held during the 1994 national AI conference, was lively and interesting. Both the theoretical and practical sides of the AI planning community were represented. Several papers contributed to the theoretical analysis of planning algorithms, and others showed the first steps toward convergence between such theoretical work and practical work on the system engineering aspects of working planners. Both the theoretical and practical sides of the AI planning community were represented, and both sides seemed to understand the other side better after the workshop. Several papers contributed further to the theoretical analysis of planning algorithms, either through frameworks for reconstructing planning algorithms or through empirical studies (Christer Backstrom, Linkoping University, Sweden; Subbarao Kambhampati, Arizona State University; Henry Kautz, AT&T Bell Labs; Craig Knoblock, University of Southern California/Information Sciences Institute [USC/ISI]; and Qiang Yang, University of Waterloo, Canada).

We describe an approach to intelligent user interfaces, based on the idea of making the computer a collaborator, and an application-independent technology for implementing such interfaces. For us, any interface that is called intelligent should at least be able to answer the six types of questions from users shown in figure 1. Being able to ask and answer these kinds of questions implies a flexible and adaptable division of labor between the human and the computer in the interaction process. Unlike most current interfaces, an intelligent user interface should be able to guide and support you when you make a mistake or if you don't know how to use the system well. What we are suggesting here is a paradigm shift. As an analogy, consider the introduction of the undo button.

My Ph.D. dissertation (Calistri 1990) extends traditional methods of plan recognition to handle situations in which agents have flawed plans. This extension involves solving two problems: determining what sorts of mistakes people make when they reason about plans and figuring out how to recognize these mistakes when they occur. I have developed a complete classification of plan-based misconceptions, which categorizes all ways that a plan can fail, and I have developed a probabilistic interpretation of these misconceptions that can be used in principle to guide a bestfirst-search algorithm. I have also developed a program called Pathfinder that embodies a practical implementation of this theory. Pathfinder is a probability-based plan-recognition system based on the A* algorithm that uses information available from a user model to guide a bestfirst search through a plan hierarchy.

Introduction In the following two subsections, we present a brief discussion of the project management problem and how the Callisto project began. The Project Management Problem Innovation is important to the continued vitality of industry. New products and changes in existing products are occurring at an increasing rate, causing product lives to decrease. In order to maintain market share, companies are forced to reduce product development time and bring their products to the market as early as possible. A major portion of development involves performing and managing many activities. For example, in hightechnology industries such as the computer industry, thousands of activities must be performed to design and build the prototype of a new product. Poor performance or management of an activity can result in critical delays. If product development time is to be reduced, better management and technical support are crucial. The Callisto project was started at the initiative of Digital ...

The methodological underpinnings of AI are slowly changing. Benchmarks, test beds, and controlled experimentation are becoming more common. Although we are optimistic that this change can solidify the science of AI, we also recognize a set of difficult issues concerning the appropriate use of this methodology. We discuss these issues as they relate to research on agent design. We survey existing test beds for agents and argue for appropriate caution in their use.