Bridging the Gap Between Planning and Scheduling

Planning research in Artificial Intelligence (AI) has often focused on problems where there are cascading levels of action choice and complex interactions between actions. In contrast, Scheduling research has focused on much larger problems where there is little action choice, but the resulting ordering problem is hard. In this paper, we give an overview of AI planning and scheduling techniques, focusing on their similarities, differences, and limitations. We also argue that many difficult practical problems lie somewhere between planning and scheduling, and that neither area has the right set of tools for solving these vexing problems. 1 The Ambitious Spacecraft Imagine a hypothetical spacecraft enroute to a distant planet. Between propulsion cycles, there are time windows when the craft can be turned for communication and scientific observations. At any given time, the spacecraft has a large set of possible scientific observations that it can perform, each having some value or priority. For each observation, the spacecraft will need to be turned towards the target and the required measurement or exposure taken. Unfortunately, turning to a target is a slow operation that may take up to 30 minutes, depending on the magnitude of the turn. As a result, the choice of experiments and the order in which they are performed has a significant impact on the duration of turns and, therefore, on how much can be accomplished. All this is further complicated by several things: • There is overlap among the capabilities of instruments, so there may be a choice to make for a given observation. Naturally, the different instruments point in different directions, so the choice of instrument influences the direction and duration of the turn. • Instruments must be calibrated before use, which requires turning to one of a number of possible calibration targets. Recalibration is not required if successive observations are made with the same instrument. • Turning uses up limited fuel and observations use power. Power is limited but renewable at a rate that depends on which direction the solar panels are facing. Given all of this, the objective is to maximize scientific return for the mission or at least to use the available time wisely. Of course, this problem is not hypothetical at all. It occurs for space probes like Deep Space One, planetary rovers like Mars Sojourner, space-based observatories like the Hubble Space Telescope, airborne observatories like KAO and SOFIA, and even automated terrestrial observatories. It is also quite similar to maintenance planning problems, where there may be a cascading set of choices for facilities, tools, and personnel, all of which affect the duration and possible ordering of various repair operations. What makes these problems particularly hard is that they are optimization problems that involve continuous time, resources, metric quantities, and a complex mixture of action choices and ordering decisions. In AI, problems involving choice of actions are often regarded as planning problems. Unfortunately, few AI planning systems could even represent the constraints in the above problem, much less perform the desired reasoning and optimization. While scheduling systems would have an easier time representing the time constraints and resources, most could not deal with the action choices in this problem. In a sense, these problems lie squarely in between planning and scheduling, and it is our contention that many important practical problems fall into this area. 1. Caelum Research Corporation 2. Research Institute for Advanced Computer Science To appear in Knowledge Engineering Review, 15(1), 2000