SEPTEMBER/OCTOBER 2004 (Vol. 6, No. 5) pp. 4-7
1521-9615/04/$31.00 © 2004 IEEE
Published by the IEEE Computer Society
Published by the IEEE Computer Society
|Intelligent Agents Peer Skyward|
|On the Grid|
|The Human Factor|
PDFs Require Adobe Acrobat
Intelligent Agents Peer Skyward
On a clear August night at the top of Hawaii's Mauna Kea, 13,400 feet above sea level, an observer peered into the night sky through the lens of the United Kingdom Infrared Telescope (UKIRT; see Figure 1 ).
Somewhere in the constellation Cygnus, the observer noticed a star that was brighter than it should have been according to the maps. It was a tell-tale flare from SS Cygni, a dwarf nova that periodically erupts with brief pulses of light and energy.
Dwarf nova flares can be unpredictable and difficult to catch, and the observer decided that UKIRT's regularly scheduled observations could wait. "Take a closer look at this," the observer told the massive telescope, issuing commands to the complex software program that controlled the instrument.
The burst of radiation from SS Cygni wasn't what made this night unusual. In fact, a team of astronomers and software engineers standing by in Hilo, at the base of the mountain, had known all along that the dwarf nova would be flaring.
The extraordinary thing was the observer: it wasn't human, but a software program, known as an intelligent agent, designed to comb through data for signs of ephemeral and little-understood events in the night sky. This agent was designed to pick out rapid changes in dwarf novae, and the observation at UKIRT last August was a test of its mettle.
If researchers at the University of Exeter and Liverpool John Moores University have their way, intelligent agents will one day be employed by astronomers all over the world. They will communicate with robotic telescopes on a worldwide network, ready to request (and, if necessary, bargain for) observation time at the first sign of a rapidly unfolding event.
The project to develop the network, known as eSTAR (eScience Telescopes for Astronomical Research), is a collaboration between the Astrophysics Research Institute at Liverpool John Moores University and the School of Physics' Astrophysics Research Group at the University of Exeter.
eSTAR is still very much in its experimental stages, but the successful test last year paved the way for the next step: deploying the software on both UKIRT and the James Clerk Maxwell Telescope, both run by the Joint Astronomy Centre in Hawaii. Over the next few years, the team hopes to add the Liverpool Telescope in La Palma, Spain; the two Faulkes telescopes in Hawaii and New South Wales, Australia; and the ASTRA (Automated Spectrophotometric Telescope Research Associates) telescope in Arizona.
If the intelligent agent works, says Tim Naylor, a professor at the University of Exeter, and a collaborator on the eSTAR team, it has the potential to go global.
"As we set the international standards for doing this, the next five years will be crucial," Naylor says. "If we get the standards right and people agree to them, then it will explode very rapidly, and a lot of people will start using it."
eSTAR could revolutionize the way astronomers study rapidly unfolding events. Currently, when an ephemeral object like a flaring dwarf nova or a gamma ray burster is spotted—often by accident—word is sent out via an email alert service. Researchers interested in that class of object must then frantically put in requests for telescope time, as the object gradually fades from the night sky.
"We currently have very little knowledge of the way things behave on short time scales, because no one can respond that quickly to changes in targets," says Andy Adamson, UKIRT's director. "At the moment, it's a mishmash of people phoning each other up on cell phones,"
With eSTAR, astronomers will not only have a network of software programs that systematically detect these events, but their own individual intelligent agent that acts as an ambassador to the network. The agent will immediately query telescopes around the globe about weather conditions and available instrumentation, bargain on behalf of its research program for observation time, collect and sift data, and deliver that data to the researcher—all without any human input.
On the Grid
The model for the eSTAR system is the computational grid: a system in which users bid for computation power distributed across many processors hooked up to a network. There is no centralized control in such a system. Rather, the various nodes in the network communicate with each other to determine which processors are currently idle and able to service a request for computing time.
With eSTAR, individual telescopes take the place of processors; the distributed resource is not data-crunching power, but observation time.
Each telescope in the eSTAR network is controlled by a software program—a discovery node—that continually makes observation data available to the network. In addition, the program juggles the scheduling of observation time, handles requests from other agents, and gives information about its capabilities and current observing conditions to the network.
"There's a definite split in the architecture between what we initially started calling discovery nodes and the agents," Naylor says. "The discovery nodes take data and put it in a scientifically usable form. On the other side are the agents, which run a science program. They're the things that decide what to observe next, and negotiate with telescopes to find out whether these objects can be observed."
The largest telescopes will be survey instruments, which handle few or no requests for observations but systematically generate massive amounts of data for the intelligent agents to pore through. (UKIRT, which is in the process of acquiring a powerful new wide-field camera, will be such an instrument.) Others will generate little data, but will be available to make observations on the fly.
The team is still considering two potential models to build. One would assign intelligent agents to individual researchers; the other would assign them to entire scientific programs.
Naylor favors the second option. "There should be one per science program. If you have a science program whose aim is to detect lots of a particular object, one agent's job will be to do that and then decide which telescope would carry out the observations."
Alasdair Allan, a professor at the University of Exeter and a colleague of Naylor's on the eSTAR team, disagrees. "I see each researcher having his or her own agent as a more architecturally elegant solution."
Allan adds that the model of one-agent-per-researcher invokes the intriguing possibility of some kind of collaboration between different agents assigned to scientists who work as a team.
"We haven't really addressed that yet," he says. "It's perfectly possible."
The Human Factor
The eSTAR team is optimistic about its system's potential. But it has some wrinkles to iron out first. Some are technical issues, involving the ability of agents to reduce and analyze data. While simple rules will suffice to detect some kinds of objects in the sky, others require increasingly sophisticated software.
"Making smart software is actually quite hard, and getting pieces of software to make decisions as well as humans from ill-defined data is actually a very hard problem," Allan says. "There are certain approaches that you can take; like neural networks and genetic algorithms and directed learning and swarm intelligence."
The intelligent agent software was built to be modular so that different rules or better intelligence could be added to it without rewriting large chunks of code. But making more intelligent agents is only part of the challenge eSTAR faces.
"The other problem is the human factor," Allan says.
"The real heart of the rule-based agents, where you code up the rules, is relatively straightforward," Naylor says. "How you get those rules out into the real world is more difficult."
The science of computational grids is still in its fledgling stages. Despite widespread interest in recent years, they remain for the most part thought experiments and research prototypes. One of the greatest barriers to developing real, functioning computational grids is the economic problem: how to charge users for access.
The difficulties are especially apparent in a system like eSTAR, where instead of a homogeneous resource like computing power, the nodes have vastly different capabilities and ever-changing conditions that affect their ability to carry out an observation.
To complicate matters, each telescope has its own system for allocating observation time. Some sell it, and some set aside a percentage of time for projects from countries that help fund the telescope. The largest telescopes (which can cost upwards of a dollar a second to run) have peer-review panels that allocate observation time based on scientific merit.
Each telescope has its own method for rescheduling time around unpredictable events, like bad weather or rapid-event researchers seeking permission to override regularly scheduled observers. Somehow, the agents on the eSTAR network will have to find ways to negotiate with a complex array of different scheduling systems.
The eSTAR team has joined forces with the Grid Markets Project ( www.lesc.ic.ac.uk/markets/) to develop economic models for grids of distributed resources. One possibility is a free-market model, in which agents bid for telescope time with either real-world money or units of some internal monetary system.
"This caused a great deal of debate at a recent conference," Allan says. "The conclusion we more or less came to is that it would have to be real money, because when you come down to it, there's nothing else that's actually valuable."
The trouble with such a model is that while it might be economically efficient, it doesn't necessarily result in the best science.
"We're not looking at a minimum amount of spend, we're looking at the maximum amount of science done," Allan says. "The thing we're trying to optimize, which is science return, is not necessarily reflected in a pure one-dimensional value like money."
Another model the eSTAR team is considering is a medieval-guild-like system, in which telescopes donate blocks of time to the network in exchange for observation time on other telescopes.
Naylor sees promise in this model. "At first, you might let the network use a little bit of time on your precious telescope, but as you realize that you're getting good quality observations in exchange, I suspect that more of your time will be taken up by the network," he says.
One thing the team won't have to do is convince the astronomical community that eSTAR is a good idea.
"Astronomers are used to new pieces of equipment coming along that can do their science better and grabbing that opportunity," Naylor says.