Space Telescope Science Institute, 3700 San Martin Dr., Baltimore MD 21218, USA, 410-338-4738, 410-338-1592
miller@stsci.edu
The HST is arguably the most complex scientific observatory every constructed and effective use of this valuable resource required novel approaches to astronomical observation and the development of advanced software systems including techniques to represent scheduling preferences and constraints, a constraint satisfaction problem (CSP) based scheduler and a rule-based planning system. This paper presents a discussion of these systems and the lessons learned from operational experience.
In addition to the constraints imposed by the proposer's scientific program, there are a large number of other constraints which must be considered. Many orbital factors exert a strong influence on scheduling: targets are occulted (blocked) by the Earth for up to 40 minutes each 95 minute orbit. The telescope cannot point too closely to the Sun, Moon or bright Earth limb. The roll orientation of the spacecraft is constrained in order to maintain correct power and thermal balance.
HST observing proposals are prepared using the Remote Proposal Submission System (RPSS) [2]. The astronomer prepares a proposal file, which is a text file in keyword-value format. The entries in this file specify the astronomical targets, exposures, instrument parameters and scientific constraints. The proposer then runs the RPSS Validation program which detects problems with the proposal file such as syntax errors, typographical errors, improper values on parameters and missing information.Validation can be performed by logging into a computer at the STScI or downloading the program via Internet and running locally. To our knowledge, RPSS was the first system of its kind for a major scientific installation (it has been in use since early 1986).
The RPSS format proposal describes the observations at a high level. The actual activities which are planned and scheduled by the downstream systems are called scheduling units and are specific realizations of the observations including details relating to spacecraft and orbital parameters and instrumental operational scenarios. The process of creating scheduling units from the proposal is called transformation and is a planning process. The STScI developed a rule-based expert system to implement transformation. When first proposed in 1984, the concept of an automated transformation of scientific proposals to implementation parameters was quite novel. Since that time, the system has demonstrated the capability to routinely perform this task and allows STScI staff to focus more attention on innovative and difficult observations. Additionally, as improved implementation strategies are devised, Transformation is quickly modified and allows us to re-transform proposals in order to benefit from these improvements. Transformation was originally implemented as a production rule-based system in OPS5, but was rewritten in Lisp as a procedural planning system [1].
Once a proposal is transformed to scheduling units, STScI staff members examine the scheduling opportunities for the proposal using the Spike system (discussed in the next section). Problems found at this stage are fixed by modifying the proposal, e.g. relaxation of observing constraints or choosing an alternate implementation strategy.
We are currently developing a second-generation RPS system which provides two major improvements over the existing system: greater insight into the planning and scheduling process and support for changes to proposals after execution has begun.
Greater insight into the planning and scheduling process is accomplished by providing the proposers with essentially the same tools as used by STScI staff, including Transformation and Spike. Graphical output will show proposers the layout of exposures and telescope activities during each orbital viewing period and the scheduling opportunities during the year, allowing them to see the implications of their choices of observing constraints, instrument parameters, etc. Proposers will also be given explicit control over the assignment of exposures to scheduling units. Previously this was determined by Transformation on the basis of a set of rules. However this was not visible to the proposer and often required several iterations with the STScI to achieve the desired groupings. Transformation will still be used to determine the detailed implementation of activities within a scheduling unit.
In addition, the proposal syntax has been enriched to allow the proposer to specify how observations should be expanded or contracted to make best use of the actual observation time (which cannot be accurately predicted more than a few months in advance).
A severe shortcoming of the current system is that once execution has begun, change to a proposal is a labor-intensive, manual process. The original ground systems were built with the assumption that most proposals would not change after submission. This has turned out to be a very poor assumption - scientific observations often require adjustment based on the results of other observations or to adjust for changes in instrument or telescope performance. Change responsiveness is being addressed in several ways. First, the overall time from proposal preparation to execution is being shortened (by about a factor of two). Second, proposal data and tools are being redesigned to be more modular so that a change to one scheduling unit or target can be processed independently of others.
A two-level, hierarchical approach has been used for HST science scheduling by dividing the problem in to long- and short-term scheduling. Long-term scheduling allocates observations over a 1-2 year interval, while short-term scheduling covers a one-week period and creates a detailed timeline of activities. Feedback from the weekly plans is used to update the long-term plan and to reschedule as needed. Long-term scheduling is performed with Spike [3] (developed at the STScI), while detailed short-term scheduling is performed with the Science Planning and Scheduling System (SPSS) which was developed by TRW and extensively modified by the STScI. Important features of Spike include:
As a scheduling tool, Spike is used to create and maintain the long-term plan. As observations are executed and proposals are created or modified, automated and manual tools in Spike are used to update the plan.
Spike was designed with generality in mind and has been adapted to about a dozen other satellite or ground-based observatories. Several of these were feasibility prototypes, but two are in flight operations: the Extreme Ultraviolet Explorer (EUVE) and the X-ray observatory ASCA. Observations with EUVE are sufficiently long (2-3 days) that a division into long- and short-term scheduling is not needed. ASCA uses a two-level hierarchy with Spike performing the long-term scheduling.
Adaptation of Spike to a new system is straightforward and largely consists of defining methods which describe mission-specific elements such as constraints. The core system which includes the constraint representation, propagation, scheduler and user interface are largely unchanged.
The adaptability of Spike was shown in another way as well - prototype short-term schedulers have been implemented for HST, ASCA and the X-ray Timing Explorer satellite. The major changes required to implement short-term scheduling included: the development of a new task which has a variable duration depending on when it is scheduled and that can be preempted (e.g. by Earth occultation or radiation belt passage); more accurate implementation of short-term constraints; a post-scheduler which adjusts task durations to utilitize small gaps remaining in the schedule.
For initial HST operations, long-term scheduling allocated each scheduling unit to a particular week. However this approach was sensitive to perturbations in the short-term schedule: If a scheduling unit could not be executed in the chosen week, this would leave a gap in the schedule which required additional effort to fill. These disruptions were sometimes caused by the fact that short-term scheduling has more information available to it and therefore can uncover problems which cannot be seen at a higher level. Another, more important, factor contributing to this was the large degree of change to HST proposals after submission - for a variety of reasons most proposals where changed after long-range planning began. The first response to this problem was to "oversubscribe" the long-range plan, that is, allocate an excess of observations to each week. In practice an oversubscription level of ~100% was necessary to ensure well-filled weekly schedules, and this made it impossible to predict with reliability when an observation would occur and required a large amount of rescheduling. We have recently developed an alternate long-term strategy which solves this problem. The long-term plan allocated each scheduling unit to a range of weeks (called a plan window). This range provides for each week an implicit oversubscription to maintain short-term efficiency, yet there is a high degree of certainty that the observation will be executed within the plan window. Our initial studies indicate that with plan windows as small as 4 weeks over 95% of the observations are executed within the plan window.
The Space Telescope Science Institute is operated by the Association of Universities for Research in Astronomy for NASA.