Scheduling Discovery in the 2020s
Eric C. Bellm, Eric B. Ford, Aaron Tohuvavohu, Michael W. Coughlin, Brett Morris, Bryan Miller, Jennifer Sobeck, Reed Riddle, Chuanfei Dong, Peter Yoachim
aa r X i v : . [ a s t r o - ph . I M ] J u l Astro2020 APC White PaperScheduling Discovery in the 2020s
Type of Activity: (cid:3) ✗ Ground Based Project (cid:3) ✗ Space Based Project (cid:3) ✗ Infrastructure Activity (cid:3) ✗ Technological Development Activity (cid:3)
State of the Profession Consideration (cid:3)
Other
Principal Author:
Name: Eric C. BellmInstitution: University of WashingtonEmail: [email protected]: 206-685-2112
Co-authors:
Eric B. Ford (Penn State)Aaron Tohuvavohu (Penn State)Michael W. Coughlin (California Institute of Technology)
Endorsers:
Brett Morris (University of Bern)Bryan Miller (Gemini Observatory)Jennifer Sobeck (University of Washington)Reed Riddle (California Institute of Technology)Chuanfei Dong (Princeton University)Peter Yoachim (University of Washington)
The 2020s will be the most data-rich decade of astronomy in history.
On the ground and inspace, powerful facilities like LSST, JWST, massive multi-object spectroscopic surveys, and awide variety of smaller robotic and queue-based telescopes will be repeatedly scanning the sky.
However, without additional effort we are unlikely to realize the full scientific potential ofour investments in these instruments.As the scale and complexity of our surveys and instruments increase, the problem ofscheduling (which observations, in what order?) becomes more critical.
First, prudencedictates maximizing the efficiency of facilities with high development costs and finite lifetimes.Second, key scientific projects—including identifying unseen populations of compact objects,understanding stellar binarity, and discovering and classifying rare classes of transients—requirecomplex history-dependent observational sequences.1 o date most scheduling of astronomical facilities has relied on very basic approaches :typically manual scheduling by human operators or simple “greedy” algorithms with basicobjective functions.
We argue that these approaches are insufficient for the scientific needs ofthe 2020s.
Thankfully, knowledge from fields such as Operations Research is beginning to percolate intoastronomy. Surveys such as LCO, ALMA, ZTF,
Swift and LSST are applying new algorithms toimprove their efficiency and quantitative scientific throughput. However, much work remains tobe done.
To maximize science in the 2020s, we must develop high-quality schedulingapproaches, implement them as open-source software, and begin linking the typicallyseparate stages of observation and data analysis.
The latter provides real-time feedbackmaximizing progress towards the scientific goal–the so-called “fifth paradigm” of science(Szalay, 2019).We provide an overview of key research directions as well as recommendations for scientists,facilities, and agencies to facilitate progress in the field. • Develop collaborations with scientists in operations research and related fields to helpadvance the state-of-the-art in astronomical scheduling and experimental design • Develop formalisms connecting scientific results to the observational sequences required • Incorporate fundamentals of experimental design into graduate curricula • Continue to organize conference tracks at regular meetings like SPIE, ADASS, etc. as wellas dedicated conferences to share ideas and new approaches. • Critically examine current scheduling and operations models and seek improvements • Develop or adopt automated scheduling approaches • Document and release scheduling tools as open-source packages to enable broad adoption • Document scheduling decisions, so that future statistical analyses can mitigate biases • Adopt a shared schedule reporting schema to allow for efficient coordination and contextualscheduling decisions The first four paradigms are observation, analytic theory, computation and simulation, and data-intensive science(Hey et al., 2009). e.g., “Artificial Intelligence in Astronomy”, For example, which observations of a star were part of a baseline planet survey and which observations were donein order to expedite publication (regardless of whether decision is made by computer or human)? e.g., via the proposed IVOA Observation Locator Table Access Protocol, .3 Funding Agencies • Formally evaluate plans for scheduling software and experimental design as part ofmission/survey proposals • Evaluate scheduling and operations models of current and forthcoming facilities • Fund interdisciplinary efforts to develop new scheduling approaches and software,particularly that with applicability beyond single missions • Fund contributions to general-purpose software libraries that schedulers and otherhigh-level packages depend on. • Recognize contributions to open-source software as a particularly high-value form ofbroader impact activity.
In the next decade, the scientific impact of a wide range of facilities can be enhanced byimprovements in scheduling.New large facilities with queue-scheduled Guest Observer programs run by time allocationcommittees will merit investments in efficient scheduling approaches appropriate to theirstaggering budgetary scale and finite lifetimes: these include JWST and the 30 m-classground-based telescopes.Large scale imaging surveys will continue to proliferate, with LSST, WFIRST, and Euclid onlythe most massive examples. Effective scheduling will be required to deliver cosmologicalconstraints with minimal (and quantifiable) bias (e.g., Bechtol et al., 2019; Capak et al., 2019;Eifler et al., 2019; Kim et al., 2019), to detect and classify rare types of transients (e.g.,Graham et al., 2019; Foley et al., 2019a), to inventory moving objects in our solar system (e.g.,Milam et al., 2019; Chanover et al., 2019), and to trace stellar populations in our Galaxy andbeyond (e.g., Kupfer et al., 2019; Rix et al., 2019; Price-Whelan et al., 2019). Smaller-aperturefacilities may choose to adjust their own observing strategies to complement larger facilities or toexplore focused niches for scientific returns too specialized for a general purpose survey.Likewise, other non-electromagnetic facilities such as ground-based gravitational waveinterferometers (e.g., LIGO) have observational selection effects (Chen et al., 2017) that, properlyexploited via coordinated scheduling, can dramatically increase the serendipitous yield ofdetected multi-messenger sources (Tohuvavohu et al., 2019).Some time-domain surveys such as ZTF and LSST will be issuing real-time alert streams withmillions of events nightly. Rapid, automated prioritization and robotic followup strategies firstapplied to gamma-ray burst followup can now be applied to a wide spectrum of time-domainevents. Such followup will be required to fully understand rare and fast-evolving time domainevents, such as kilonovae, “orphan” afterglows of gamma-ray bursts, rare types of stellarvariables, and more.Some of this followup will be undertaken by distributed networks of fully robotic telescopes (e.g.,3he Las Cumbres Observatory) which may be scheduled as a single entity (Lampoudi et al., 2015).However, an even greater opportunity and challenge is to realize the potential of heterogeneoustelescope networks (e.g., Hessman, 2014) to stitch together disparate facilities in a uniform way.This can be useful for a few purposes. For example, using telescopes in both hemispheres can beimportant due to significant size of the sky localizations associated with short gamma-ray burstsand gravitational-wave sources. Often localizations have probability that is not fully accessiblefrom a single location. It can also be useful to reimage the same localizations with multiplesystems, both in different filters to measure an object’s color, as well at different times todifferentiate between real transients and asteroids, as well as potentially measuring a change inluminosity. Coordinated scheduling between systems with significant differences betweentelescope setups, including their placement on the Earth, and their instrument configurations,including field of view, filters, typical exposure times, and limiting magnitudes, is desired.Existing networks have begun to implement this network-level capability in open-sourcecodebases (e.g., Coughlin et al., 2018) . The Astronomical Event Observatory Network (AEON )represents a current and more general step towards connecting multiple observatories through acommon interface (Saunders et al., 2018, and see the Miller et al. 2019 APC white paper“Infrastructure and Strategies for MMA and Time Domain Follow-Up”), with a goal of enablingnetwork-level scheduling of both imaging and spectroscopy for multiple teams on disparateground-based facilities around the world.Multi-wavelength and multi-team followup efforts, such as electromagnetic followup ofgravitational wave triggers (e.g., Foley et al., 2019b; Sathyaprakash et al., 2019), present anadditional layer of complexity. Coordinating disparate and perhaps competing PI teams using keyshared facilities, each with their own TOO policies and scheduling constraints, is as much apolitical problem as an algorithmic one. But greater sharing of observations planned andundertaken could enable individual teams to pursue quantitatively rigorous approaches.Massive multi-object spectroscopic surveys (DESI, PFS, SDSS-V, WEAVE, 4MOST, and more)will need to dynamically adapt their targeting strategy to account for changing observingconditions. The SDSS-V survey in particular has a strong time-domain component which willrequire effective scheduling (e.g., Rix et al., 2019; Kollmeier et al., 2019; Shen et al., 2019). Withappropriate investment they could also opportunistically allocate spare fibers dynamically tofollow up targets identified by time-domain imaging surveys.Massive radial velocity surveys for exoplanets will demand rigorous scheduling approaches toanswer key population questions (e.g., Ford et al., 2019; Bryson et al., 2019).And finally, individual investigators with classically-scheduled nights will continue to seek tomake the most effective use of the time they have available.Across all of these diverse fields, improved scheduling methods can provide a range of benefits,including: https://github.com/mcoughlin/gwemopt http://ast.noao.edu/data/aeon Greater scientific productivity and reduced overheads • Greater transparency and reproduceability • Improved ability to rigorously simulate selection functions • Potentially lower operations costs if replacing labor-intensive manual schedulingapproaches (e.g., manually producing several versions of queue schedules for potentiallydifferent observing condition bands, most of which won’t occur). • The ability to dynamically respond to failures, changing observing conditions, and target ofopportunity requests by repeatedly and regularly re-solving the scheduling problem (e.g.,Lampoudi et al., 2015).However, implementing state-of-the-art scheduling methods is currently not without itschallenges: • Initial development of new scheduling approaches requires specialized skills which arequite rare in astronomy. • Changes to operations models and program prioritization may encounter user resistanceunless the benefits (and any tradeoffs) are made clear.The costs of implementing such scheduling approaches would be lowered if there werewidely-available open-source scheduling software for the major classes of astronomicalscheduling problems. The need to develop such algorithms and software forms the core of ourrecommendations. Current surveys are continuing to develop innovative scheduling approaches. These include bothadvances in the metrics or objective functions that surveys seek to maximize as well as newoptimization algorithms.
Many facilities on the ground and in space operate largely in a queue-based guest observer mode.Typically these observatories have disparate instruments and/or configurations, and the observingprograms are selected on a periodic basis by a time allocation committee. Such facilities arewell-served by optimization metrics that assess whether the proposals with higher TAC rank wereexecuted in preference to those with lower rank, perhaps weighted by observing conditions.Lampoudi et al. (2015) and Solar et al. (2016) describe TAC-priority objective functions for theLas Cumbres Observatory robotic telescope network and ALMA, respectively.Given their inherent multiplexing, large-scale imaging surveys often (but not exclusively) pursuemultiple science goals simultaneously. They are accordingly well-served by objective functionsthat capture the rate at which they survey spatial volume (e.g., Bellm, 2016; Bellm et al., 2019) or See the E. Tollerud et al. APC whitepaper, “Sustaining Community-Driven Software for Astronomy in the 2020s,”for more on the scientific importance of open-source development and recommendations for sustaining it.
While the science metric to optimize will be specific to the survey or facility, in many cases thescheduling algorithm may be decoupled from the metric chosen. While astronomical schedulingproblems have clear analogues in the optimization literature, most are different enough fromstandard textbook problems to prevent easy application of known solutions. For example, theproblem of minimizing slew time when pointing a telescope can be cast as a Travelling SalesmanProblem. However, for ground-based telescopes the earth’s rotation makes the time to slewbetween two points time-dependent. Moreover, the value of observing the fields at a specific time(the science metric) also is likely to change with time, for example as the airmass of the objectchanges (cf. Bellm et al., 2019).Likewise, for space-based telescopes the slew-time between two points varies with timedepending on many factors, including the exact dynamic Attitude Control System constraints ofthe spacecraft. Analogously, the value of observing fields at a specific time can change dependingon the distance of the object from the Earth-limb at that time, for example (cf. Tohuvavohu, 2017).Because many optimization problems have only heuristic solutions, the optimization literature islarge, and many approaches have been explored by different astronomical projects (for a review,see Solar et al., 2016, and references therein). We highlight a handful of the wide range ofapproaches reported in the literature. In many cases hybrid approaches that apply differentalgorithms for short- and long-term scheduling are used (e.g., Colom´e et al., 2010;Garcia-Piquer et al., 2014; Wetter et al., 2015).“Greedy” or local search-based scheduling is quite common. The known shortcomings of greedysearches have led some authors to propose simple extensions that take into accountastronomy-specific features like targets rising and setting (e.g., Denny, 2006; Rana et al., 2016).Others have adopted optimization approaches that heuristically attempt to identify better solutionsthan the current local optimum, including simulated annealing, genetic algorithms, Tabu Search,and Ant Colony Optimization (e.g., Colome et al., 2012; Garcia-Piquer et al., 2014).In some cases generating a satisfactory observing sequence that avoids certain instrumentallimitations is more important than maximizing an objective function. This suggests casting thescheduling problem in Constraint Satisfaction terms. A major example is the SPIKE system(Johnston & Miller, 1994), which has been used by the Hubble Space Telescope, Spitzer,Chandra, and the ground-based VLT. A related approach, TAKO (Timeline Assembler KeywordOriented), has been utilized by
Swift (2004–2018),
Fermi , Suzaku and others. The
Swift
Scheduler(Tohuvavohu, 2017) and the open-source astroplan module (Morris et al., 2018) also employ6elated approaches, though the
Swift
Scheduler combines this approach with a dynamic, fuzzyweighted priority scheme (Luo et al., 2003; Verfaillie & Jussien, 2005) and a flexible suite ofoptimization (objective) functions that yield a framework with generalities and automationcapabilities more similar to Integer Linear Programming approaches.Integer Linear Programming (ILP)-based approaches have been adopted by Las CumbresObservatory (Lampoudi et al., 2015), ALMA (Solar et al., 2016), and ZTF (Bellm et al., 2019).These provide a more general framework for optimizing objective functions than constraintsatisfaction while still enabling rigorous handling of constraints. Powerful commercial librariesare readily available . Casting the scheduling problem in ILP terms enables rapid, regularreoptimization to update the schedule to respond to the success or failure of observations,changing weather conditions, etc.Finally, some schedulers are based on artificial intelligence “agents” that are pre-trained to makedecisions about observation sequences and then run on that trained model. Reinforcementlearning approaches such as that of the LSST scheduler (e.g., Naghib et al., 2019) fall into thiscategory. Reinforcement learning using deep neural networks has generated a great deal ofpopular press thanks to Google’s AlphaGo (Silver et al., 2016, 2017, 2018).This profusion of solutions highlights the potential for breakthrough new approaches, but it alsoraises questions. How can we compare the performance of disparate systems applied to differentscheduling problems? What are the performance and operational tradeoffs between differentalgorithmic approaches? Are there specific scheduling strategies that are provably superior forcertain classes of problems found in astronomy? In addition to the exciting new astronomical facilities which motivate us to seek improvedscheduling ( § . Cost per CPU core should continue todecrease, aiding optimization approaches which can be parallelized. Improved tooling for portingcode to run on GPUs can yield even larger performance gains, and we are beginning to seeapplication-specific integrated circuits such as Tensor Processing Units (TPUs) that arespecifically developed for optimization workflows.In conjunction with further algorithmic work ( § e.g., Gurobi, IBM CPLEX cf. the A. Smith et al. APC whitepaper “Astronomy Should be in the Clouds.”
7e take next to most improve our constraint on the dark energy figure of merit?”
We prioritize five key areas for further development in the 2020s:First, the development of a more general taxonomy of astronomical scheduling problems willilluminate mathematical commonalities and differences and guide algorithm selection.Second, refinement of science-based metrics for optimization will enable surveys and facilities tobetter quantify their progress towards their scientific goals.Third, development of more powerful scheduling algorithms will improve the overall throughputof existing facilities at modest cost.Fourth, development of well-tested, well-documented open source implementations of schedulingalgorithms will encourage reuse in the major problem domains, lowering the barrier to entry. The astropy -affiliated astroplan package provides one potential framework to whichscheduling modules could be contributed.Finally, connecting observations to data reduction and analysis will enable real-time optimizationof a survey’s ultimate science goals. For example, an exoplanet RV survey might ask, “How doesthis observation improve my knowledge of the mass of a planet?” (Ford, 2008) or even better,“How does the observation improve my knowledge of the mass distribution of 2–4 R ⊕ planetswith orbital periods of 30–300 d?” Research in the next decade will investigate how to make this“fifth paradigm” of science (Szalay, 2019) computationally feasible while preserving secondaryscience objectives and serendipitous discoveries. References
Bechtol, K., Drlica-Wagner, A., Abazajian, K. N., et al. 2019, in BAAS, Vol. 51, 207Bellm, E. C. 2016, PASP, 128, 084501Bellm, E. C., Kulkarni, S. R., Barlow, T., et al. 2019, PASP, 131, 068003,doi:
Bryson, S., Bennett, D., Gaudi, S., et al. 2019, in BAAS, Vol. 51, 443Capak, P., Scolnic, D., & Davidzon, I. 2019, in BAAS, Vol. 51, 470Chanover, N., Wong, M. H., Greathouse, T., et al. 2019, in BAAS, Vol. 51, 340Chen, H.-Y., Essick, R., Vitale, S., Holz, D. E., & Katsavounidis, E. 2017, The AstrophysicalJournal, 835, 31, doi:
Colom´e, J., Casteels, K., Ribas, I., & Francisco, X. 2010, in Proc. SPIE, Vol. 7740, Software andCyberinfrastructure for Astronomy, 77403KColome, J., Colomer, P., Gu`ardia, J., et al. 2012, in Proc. SPIE, Vol. 8448, ObservatoryOperations: Strategies, Processes, and Systems IV, 84481LCoughlin, M. W., Tao, D., Chan, M. L., et al. 2018, Monthly Notices of the Royal Astronomical8ociety, 478, 692, doi:
Denny, R. B. 2006, Dispatch Scheduling of Automated Telescopes, http://scheduler.dc3.com/Scheduler2006.pdf
Eifler, T., Simet, M., Hirata, C., et al. 2019, in BAAS, Vol. 51, 418Foley, R., Bloom, J. S., Cenko, S. B., et al. 2019a, in BAAS, Vol. 51, 305Foley, R., Alexander, K. D., Andreoni, I., et al. 2019b, in BAAS, Vol. 51, 295Ford, E., Bedell, M., Ciardi, D. R., et al. 2019, Astro2020: Decadal Survey on Astronomy andAstrophysics, 2020, 466Ford, E. B. 2008, AJ, 135, 1008, doi:
Garcia-Piquer, A., Gu`ardia, J., Colom´e, J., et al. 2014, in Proc. SPIE, Vol. 9152, Software andCyberinfrastructure for Astronomy III, 915221Graham, M., Milisavljevic, D., Rest, A., et al. 2019, in BAAS, Vol. 51, 339Hessman, F. V. 2014, in The Third Hot-wiring the Transient Universe Workshop, ed. P. R.Wozniak, M. J. Graham, A. A. Mahabal, & R. Seaman, 95–101Hey, T., Tansley, S., &Tolle, K. 2009, The Fourth Paradigm: Data-Intensive Scientific Discovery (Microsoft Research).
Johnston, M. D., & Miller, G. E. 1994, in Intelligent Scheduling, ed. M. Zweben & M. S. Fox(Morgan Kaufmann Publishers), 391Kim, A., Aldering, G., Antilogus, P., et al. 2019, in BAAS, Vol. 51, 140Kollmeier, J., Fuller, J., Gaensicke, B., et al. 2019, in BAAS, Vol. 51, 503Kupfer, T., Kilic, M., Maccarone, T., et al. 2019, in BAAS, Vol. 51, 188Lampoudi, S., Saunders, E., & Eastman, J. 2015, ArXiv e-prints. https://arxiv.org/abs/1503.07170
Loredo, T. J. 2004, in American Institute of Physics Conference Series, Vol. 707, BayesianInference and Maximum Entropy Methods in Science and Engineering, ed. G. J. Erickson &Y. Zhai, 330–346Luo, X., man Lee, J. H., fung Leung, H., & Jennings, N. R. 2003, Fuzzy Sets and Systems, 136,151 , doi: https://doi.org/10.1016/S0165-0114(02)00385-8
Milam, S., Hammel, H., Bauer, J. M., et al. 2019, in BAAS, Vol. 51, 327Morris, B. M., Tollerud, E., Sip˝ocz, B., et al. 2018, AJ, 155, 128,doi:
Naghib, E., Yoachim, P., Vanderbei, R. J., Connolly, A. J., & Jones, R. L. 2019, AJ, 157, 151,doi:
Price-Whelan, A., Breivik, K., D’Orazio, D., et al. 2019, in BAAS, Vol. 51, 206Rana, J., Singhal, A., Gadre, B., Bhalerao, V., & Bose, S. 2016, ArXiv e-printsRix, H.-W., Ting, Y.-S., Sippel, A., et al. 2019, in BAAS, Vol. 51, 104Sathyaprakash, B., Bailes, M., Kasliwal, M. M., et al. 2019, in BAAS, Vol. 51, 276Saunders, E. S., Bowman, M., Boroson, T., & Street, R. A. 2018, in Society of Photo-OpticalInstrumentation Engineers (SPIE) Conference Series, Vol. 10704, Proc. SPIE, 107040ZShen, Y., Anderson, S., Berger, E., et al. 2019, in BAAS, Vol. 51, 274Silver, D., Huang, A., Maddison, C. J., et al. 2016, Nature, 529, 484,doi:
Silver, D., Schrittwieser, J., Simonyan, K., et al. 2017, Nature, 550, 354,doi:
Solar, M., Michelon, P., Avarias, J., & Garces, M. 2016, Astronomy and Computing, 15, 90Szalay, A. 2019, in American Astronomical Society Meeting Abstracts, Vol. 233, AmericanAstronomical Society Meeting Abstracts https://arxiv.org/abs/1712.08111
Tohuvavohu, A., Evans, P., Kennea, J., & Cenko, S. B. 2019, in American Astronomical SocietyMeeting Abstracts, Vol. 233, American Astronomical Society Meeting Abstracts 10.1007/s10601-005-2239-9