{"id":265938,"date":"2014-06-03T10:54:53","date_gmt":"2014-06-03T17:54:53","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/?p=265938"},"modified":"2017-06-25T15:03:46","modified_gmt":"2017-06-25T22:03:46","slug":"ai-takes-skies","status":"publish","type":"post","link":"https:\/\/www.microsoft.com\/en-us\/research\/blog\/ai-takes-skies\/","title":{"rendered":"AI Takes to the Skies"},"content":{"rendered":"

As do other recreational pilots, Ashish Kapoor (opens in new tab)<\/span><\/a> learned during flight training that he shouldn\u2019t count on the accuracy of wind forecasts. The best available forecasts in the United States\u2014from the federal government\u2019s Winds Aloft program\u2014have been based largely on data from instrumented weather balloons released twice a day, providing forecasts for 176 stations across the United States.<\/p>\n

\u201cThey would tell you, \u2018Winds Aloft is often not accurate, so you have to take that into account when you make your flight plan,\u2019\u201d says Kapoor, a Microsoft researcher who specializes in machine learning and decision-making. Pilots learn to plan for longer flight times and greater fuel consumption, he says, in case the actual wind conditions are not as expected.<\/p>\n

In our data-driven, sensor-filled world, that lack of precision motivated an effort to devise a better solution.<\/p>\n

\"Ashish

Ashish Kapoor (left) and Eric Horvitz<\/p><\/div>\n

Kapoor and his colleague Eric Horvitz (opens in new tab)<\/span><\/a>, a Microsoft distinguished scientist and managing director of Microsoft Research Redmond (opens in new tab)<\/span><\/a>, began looking for such a solution\u2014specifically, one that would not involve installing new sensors or other equipment. They focused on using aircraft in flight as a possible source of data.<\/p>\n

It turns out that data about the tens of thousands of flights over the United States each day are publicly available from the Federal Aviation Administration (FAA). That means people can view any commercial aircraft\u2019s flight plan\u2014the planned airspeed (fixed cruising speed relative to the air mass), altitude, distance, and route\u2014as well as the observed groundspeed (speed relative to the ground), altitude, longitude, and latitude at any moment during the flight, with only a small time lag.<\/p>\n

Live access: Real-time Windflow map of the continental United States<\/strong> (opens in new tab)<\/span><\/a><\/p>\n

\u201cOur research question was: Could we take the information that was already available and use it to predict wind conditions without needing any additional infrastructure?\u201d Kapoor says. In other words, could airplanes in flight be employed as a vast sensor network to determine atmospheric conditions? Could data available today be used to infer winds on a large scale without special plane-based wind sensors and new infrastructure to access and combine signals from planes?<\/p>\n

Because winds will change an aircraft\u2019s flight path and affect the groundspeed, the researchers determined that they could estimate the wind conditions mathematically.<\/p>\n

That sounds straightforward enough\u2014if no wind is present, the groundspeed will match the airspeed and the plane will fly along its intended course. If you know the groundspeed and the airspeed, along with where the plane is heading and its actual course over the ground, you can calculate the wind speed and direction.<\/p>\n

But one key piece of data was unavailable: The data provided by the FAA does not include information about where a plane is headed.<\/p>\n

Solving this conundrum involved a creative leap\u2014what Kapoor calls their \u201cbiggest intellectual aha moment.\u201d They realized that winds exhibit a special property known as spatial regularity\u2014that is, nearby airplanes are likely to encounter similar winds. They then figured out a way to exploit this property by taking observations from nearby wind stations, combining it with the FAA data, and then using a probabilistic model to infer the wind velocity. In essence, they came up with a way to solve a puzzle despite many missing pieces.<\/p>\n

Comparing Predictions with Ground Truth<\/h2>\n

To help with testing the accuracy of their model, which they call Windflow (opens in new tab)<\/span><\/a>, Kapoor and Horvitz\u2014along with two high school researchers, Spencer Laube and Horvitz\u2019s son, Zachary, both students at the Seattle Academy of Arts and Sciences\u2014launched a helium-filled high-altitude balloon in eastern Washington state in June 2013. Equipped with a GPS device, barometric and temperature sensors, and an onboard computer, the balloon was carried by the wind, reaching an altitude of 95,000 feet.<\/p>\n

\"Zachary

Zachary Horvitz (left) and Spencer Laube prepare to launch an instrumented high-altitude balloon.<\/p><\/div>\n

Before launch, they estimated the endpoint of the balloon\u2019s journey\u2014when it would reach maximum altitude and rupture\u2014using their own model, as well as the Winds Aloft model from the National Oceanic and Atmospheric Administration (NOAA). The results were striking: The Winds Aloft prediction was off by 56.2 miles, in an easterly direction, while the researchers\u2019 prediction was off by only 11.6 miles, in the same direction.<\/p>\n

A small aircraft would spend much less time traveling the same route and climbing through those altitudes, but the NOAA error is \u201ca pretty big cumulative error in terms of winds across altitude profiles,\u201d Kapoor says.<\/p>\n

It could certainly be significant for high-altitude balloons deployed for research purposes, which often carry expensive instrumentation that parachutes to the ground when the balloon bursts at a certain height.<\/p>\n

\u201cThe error,\u201d Kapoor says, \u201ccould make a difference in the ability to recover the balloon\u2019s payload.\u201d<\/p>\n

In addition to publishing their findings in a paper (opens in new tab)<\/span><\/a>, Kapoor and Horvitz have made their predictive model publicly available on the Microsoft Azure (opens in new tab)<\/span><\/a> cloud service. The service graphically depicts the Windflow model, the Winds Aloft model, and the differences between them on a map of the continental United States.<\/p>\n

The views are computed every three hours during the daytime, using a subset of aircraft rather than all of the aircraft in flight.<\/p>\n

\u201cIt turns out that the biggest obstacle was getting the data\u2014we have to pay for each piece of data from each aircraft,\u201d Kapoor says. This is because the only practical way for the researchers to access the FAA data\u2014for now\u2014is through a commercial website that charges for each ping to an aircraft.<\/p>\n

\u201cThe cost becomes prohibitive,\u201d says Kapoor. \u201cWe had a budget, so we had to decide which planes to track.\u201d<\/p>\n

Their budget, which is enough to track 100 planes, led Kapoor and Horvitz into another area of research that also is detailed in their paper\u2014computing the \u201cexpected value of information\u201d that each plane might contribute to determine which 100 planes to sample at any given time to yield the most valuable data for wind prediction over the entire continent.<\/p>\n

\"every

Every airplane in flight can act as a sensor to help estimate large-scale wind conditions.<\/p><\/div>\n

These days, when Kapoor is out on a weekend flight\u2014usually in a two-seater or a four-seater\u2014he says he uses the Windflow predictions in addition to NOAA forecasts for his flight plans:<\/p>\n

\u201cI look at the Windflow data on my phone.\u201d<\/p>\n

Although most of the Windflow data is for higher altitudes where commercial aircraft generally fly, the predictions go as low as 6,000 feet, which can be useful to recreational pilots such as Kapoor. Not only does Windflow provide finer resolution of wind directions than Winds Aloft, but Kapoor also has found that its wind-velocity predictions are more accurate.<\/p>\n

Kapoor and Horvitz foresee an array of applications for the Windflow methodology.<\/p>\n

\u201cThe ongoing inferences can be used to identify ideal routes for aircraft, enabling faster trips that use less fuel and output less carbon dioxide,\u201d Horvitz says. \u201cOther applications include using the finer-grained wind reports for guiding and predicting the trajectories of gliders and balloons, whether they are being used for communications, sensing, or recreation.\u201d<\/p>\n

The results also can be used to improve understanding of phenomena such as turbulence and larger-scale weather processes, as well as for storm tracking\u2014an area in which, Kapoor notes, current predictive methodologies are particularly unreliable.<\/p>\n

In the Flow<\/h2>\n

Windflow is the latest in a series of efforts at Microsoft Research to use machine learning and reasoning to create new kinds of services.<\/p>\n

Smartflow (opens in new tab)<\/span><\/a>, created more than a decade ago, uses highway-flow data, accident reports, weather, and major local events to predict where traffic bottlenecks will form and how long they will last.\u00a0The system even performs surprise forecasting (opens in new tab)<\/span><\/a>\u2014using machine learning about surprising situations over many years to predict when people likely will be surprised by atypical jams or unexpected wide-open flows and alerting drivers before the flows or jams happen.<\/p>\n

That project led to an even more intensive effort called Clearflow (opens in new tab)<\/span><\/a>, which uses large quantities of highway and GPS data, along with the topology or structure of the traffic network, to predict flows on all streets in a greater metropolitan area. The inferences from Clearflow have been used to generate traffic-sensitive directions (opens in new tab)<\/span><\/a> in Bing Maps (opens in new tab)<\/span><\/a> since 2008, providing enhanced routing information for cities throughout North America.<\/p>\n

\u201cWindflow,\u201d Horvitz says, \u201cis a great example of where things are headed with using machine intelligence to perform sensing, learning, and reasoning about large-scale phenomena like winds and weather\u2014and to stream the insights down to people who can make use of the inferences in their decision-making.<\/p>\n

\u201cI\u2019m very excited about the results and directions with Windflow.\u201d<\/p>\n","protected":false},"excerpt":{"rendered":"

As do other recreational pilots, Ashish Kapoor learned during flight training that he shouldn\u2019t count on the accuracy of wind forecasts. The best available forecasts in the United States\u2014from the federal government\u2019s Winds Aloft program\u2014have been based largely on data from instrumented weather balloons released twice a day, providing forecasts for 176 stations across the […]<\/p>\n","protected":false},"author":39507,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"footnotes":""},"categories":[194466,194467,205399,194474,194478,194455,194483],"tags":[208917,208929,208923,208935,208914,208920,208926,208911,208932],"research-area":[13561,13556,13563,198583,13546],"msr-region":[],"msr-event-type":[],"msr-locale":[268875],"msr-post-option":[],"msr-impact-theme":[],"msr-promo-type":[],"msr-podcast-series":[],"class_list":["post-265938","post","type-post","status-publish","format-standard","hentry","category-algorithms","category-artifical-intelligence","category-azure","category-data-visulalization","category-ecology-and-environment","category-machine-learning","category-mathematics","tag-clearflow","tag-federal-aviation-administration-faa","tag-flight-plan","tag-national-oceanic-and-atmospheric-administration-noaa","tag-smartflow","tag-surprise-forecasting","tag-wind-forecasts","tag-windflow","tag-winds-aloft","msr-research-area-algorithms","msr-research-area-artificial-intelligence","msr-research-area-data-platform-analytics","msr-research-area-ecology-environment","msr-research-area-computational-sciences-mathematics","msr-locale-en_us"],"msr_event_details":{"start":"","end":"","location":""},"podcast_url":"","podcast_episode":"","msr_research_lab":[199565],"msr_impact_theme":[],"related-publications":[],"related-downloads":[],"related-videos":[],"related-academic-programs":[],"related-groups":[],"related-projects":[256167],"related-events":[],"related-researchers":[],"msr_type":"Post","byline":"","formattedDate":"June 3, 2014","formattedExcerpt":"As do other recreational pilots, Ashish Kapoor learned during flight training that he shouldn\u2019t count on the accuracy of wind forecasts. The best available forecasts in the United States\u2014from the federal government\u2019s Winds Aloft program\u2014have been based largely on data from instrumented weather balloons released…","locale":{"slug":"en_us","name":"English","native":"","english":"English"},"_links":{"self":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/posts\/265938"}],"collection":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/users\/39507"}],"replies":[{"embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/comments?post=265938"}],"version-history":[{"count":6,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/posts\/265938\/revisions"}],"predecessor-version":[{"id":393461,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/posts\/265938\/revisions\/393461"}],"wp:attachment":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media?parent=265938"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/categories?post=265938"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/tags?post=265938"},{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=265938"},{"taxonomy":"msr-region","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-region?post=265938"},{"taxonomy":"msr-event-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-event-type?post=265938"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=265938"},{"taxonomy":"msr-post-option","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-post-option?post=265938"},{"taxonomy":"msr-impact-theme","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-impact-theme?post=265938"},{"taxonomy":"msr-promo-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-promo-type?post=265938"},{"taxonomy":"msr-podcast-series","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-podcast-series?post=265938"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}