Use this URL to cite or link to this record in EThOS: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.570885
Title: Adaptive optimal telescope scheduling
Author: Fraser, Stephen Nicholas
Awarding Body: Liverpool John Moores University
Current Institution: Liverpool John Moores University
Date of Award: 2012
Availability of Full Text:
Access through EThOS:
Abstract:
Astronomical telescopes operate in a dynamic and uncertain operational environment. Time is often oversubscribed with programs competing for available slots and the best observing conditions. In order to achieve a high scientific return a scheduler must be able to create a globally optimal observing schedule. In dynamic environments any offline or static schedule will become rapidly out-of- date, thus iterative repair based or dynamic despatch scheduling is the preferred option being able to respond quickly to such changes. Dynamic despatch suffers from myopism. Long term global aims are usurped in favor of local optimization, namely what is best to do right now. Using data collected from external and embedded instrumentation I investigate and characterize the scheduler's operating environment and investigate tech- niques for short term prediction. I investigate metrics for characterizing the value of schedules both locally (at the decision point) and over longer horizons. Using this information an extensible software architecture is designed incorporating a simulation framework to allow a variety of schedulers to be implemented. Experiments are performed using the scheduler component architecture and simulation framework to determine the effects on schedule quality of environmental stability, disruptive events and reliability of prediction under a range of load conditions. Ultimately these show that where conditions are relatively stable there is an advantage to using longer look-ahead horizons but that where conditions are unstable or disruptive a basic despatch scheduler achieves results as good or better. A look-ahead scheduler with a horizon of 4 hours can give an improvement of upto 26% over a simple despatch scheduler when seeing remains stable over a period of up to 6 hours. When the loading is high, further experiments show that look-ahead schedulers can achieve better quality than despatching. As load (measured by con- tention) is increased from 1 to 25 (a reasonable range), performance of a despatcher increases by upto 17% while for a look-ahead scheduler with 4 hour horizon it can increase by upto 23%. In the experiments the look-ahead scheduler has an advantage of between 10% and 14% over the despatcher over this range of loads. Under disruptive conditions (breaks in the normal execution of the schedule) due to weather or mechanical problems, it was found that a large number of small disruptions has a more negative effect on schedule quality than a few long events. The despatcher was unaffected by these disruptions while the look-ahead sched- uler under-performed by as much as 11 % relative to the despatcher and upto 22% relative to its own performance under non-disruptive conditions which were typ- ically 14% better than the despatcher. When new observations are being added to the pool during schedule execution it was found that a look-ahead scheduler could loose out on the potential gains of adding high quality observations if the lead-time for these observations was small compared to the scheduler's look-ahead horizon. Typically the highest gain occurs where the lead-time is of the order of the look-ahead horizon. The amount of gain also decreases for longer horizons. This loss can be up to 40% for long (4 hour) horizon and as little as a few percent for a horizon of 30 minutes. For very short lead-time observations, the loss of potential reward can increases for longer horizons. For a 4 hour horizon as much as 80% of the potential gain can be lost. Under such conditions, the addition of a sequestration mechanism by means of which newly arriving high quality urgent observations can be added into an executing schedule are shown to improve the performance of look-ahead schedulers by upto 8.8% with a look-ahead horizon of 4 hours but with little effect for shorter horizons. Further experiments have shown that when determining the length of look- ahead horizon, if we either over or under-estimate the duration of stable con- ditions we obtain poorer results. When the stability length is severely under- estimated a loss of between 5% and 17% in quality was found to occur. The optimum situation appears to be when the horizon is of order the stability length. Where the stability length is over-estimated by a factor 2, a loss of 5% to 10% can occur.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.570885  DOI: Not available
Share: