Workshop overview

Many of the significant advances of the machine learning community have come from tackling challenging, real world problems of large practical and economic importance. The application of ML techniques has often produced substantial savings in human time, energy, and money. The inherent flexibility in ML techniques is particularly relevant for space missions, where multi-million dollar hardware must operate without a human overseer, in sometimes unforeseeable situations.

Space environments impose several stringent resource and time constraints. Remote space missions must cope with very limited bandwidth and, often, response time delays too large to allow direct human teleoperation. Currently, limited bandwidth is often dealt with by shutting down science instruments entirely until the current batch of data has been downlinked. Likewise, if a pre-loaded command sequence fails for some reason, a remote rover is often programmed to shut down and sit idle until the next direct Earth communications opportunity.

In contrast, methods designed to intelligently filter observations, compress redundant information, and devote more time and bandwidth to novel, unusual events, could greatly enhance the efficiency and science return of such missions. Reinforcement learning and other techniques could be useful for adapting to changing environments, adjusting science objectives, making navigational decisions, and generally enhancing the autonomy of our remote, robotic delegates.

Despite progress in developing applicable ML techniques, adoption and integration into fielded remote space missions remains a challenge. Scientists and mission PIs are reluctant to entrust critical mission components, such as science target identification and vehicle control, to methods that take decision making out of human hands. Similar concerns arise in any situation where failure brings with it an unacceptably high cost. This hesitation, while understandable, is at odds with our desire to overcome the sometimes formidable communication delays, to make the best use of severely limited bandwidth, and to improve the science return of such missions.

In recognition of the gap between the development of ML methods and their adoption in fielded space systems, we convene this workshop as a forum where we can address the following critical questions:

• How can we design algorithms that can train for a long time under controlled situations, but must work almost perfectly in a remote, autonomous setting?
• How can ML techniques be tested so as to convince someone outside the field that they are reliable, robust, and effective for real space systems? What are the best analogue problems and situations, here on Earth, for the development and study of applicable ML techniques?
• Are there specific, possibly novel, metrics and methodologies for evaluation that would be most appropriate for these problems?
• What ML algorithms drawn from other domains (e.g., tasks with a high cost of failure) are applicable to the problems faced by fielded space missions?
• Can we provide formal performance guarantees for ML algorithms in the constrained and sometimes hostile environments in which remote space systems will exist?
• How can we strengthen connections between ML researchers and the people making operational decisions for space missions?

We are interested in discussing both social and political barriers to adoption of ML methods in space applications, and technological solutions that would help erode those barriers. We will also welcome contributions from the entire spectrum of associated fundamental, methodological, and practical research.

Submissions

Because we wish to encourage discussion and maintain a small group atmosphere, attendance will be limited to invitation only. We welcome your innovative, controversial, yet well reasoned ideas. We want to spark productive debates. Those wishing to participate in this workshop should submit one (or more) of the following:

1. A 3-5 page paper on technical challenges/solutions for ML in space missions (operational constraints, methods best suited to those constraints, best Earth-analog test beds, etc.)
2. A 3-5 page paper on social challenges/solutions for ML in space missions (adoption, reliability, how to convince the mission decision-makers, etc.)
3. A 3-5 page paper on opportunities for ML in space missions (what planned missions stand to benefit most, who are the people to convince, etc.)

   For those not wishing to contribute a technical paper, there is also a fourth option:

4. A one page statement of your relevant interests and how you can contribute to the workshop. This will not be published and will only be used to select attendees from those who do not wish to contribute a technical paper. This category has a later submission date than the first three, but invitation preference will be given to technical submissions and to early submissions of relevant interests.

In addition, submitting authors must include (as a separate file), up to 500 words of ASCII text describing their biography, relevant experience, and contact information. The purpose of these biographies is to assist attendees in getting to know others with relevant or complementary interests. They will appear in the published workshop notes. We ask that authors not include lists of publications, but we will be happy to accept URLs pointing to personal research pages.

• Technical submissions due: May 5, 2003
• Notification of acceptance: May 27, 2003
• Camera ready copies of both bios and technical submissions: June 13, 2003
• Workshop date: August 21, 2003

Last changed Tue, Aug 12, 2003 20:46:33 .