back to article Aardvark beats groundhogs and supercomputers in weather forecasting

Aardvark, a novel machine learning-based weather prediction system, teases a future where supercomputers are optional for forecasting - but don't pull the plug just yet. Academics affiliated with the Alan Turing Institute in the UK and other institutions claim they have developed a weather prediction model that can be trained …

  1. IanRS

    Simple model

    Weather(tomorrow) = Weather(today)

    Accurate about 80% of the time, but sometimes gets it drastically wrong.

    1. LybsterRoy Silver badge

      Re: Simple model

      and then there is the classic "LOOK OUT THE WINDOW" model

      1. Giles C Silver badge
        Joke

        Re: Simple model

        Or get the famous weather forecasting rock…

        And Train the AI on these rules

        If the rock is wet, it's raining.

        If the rock is swinging, the wind is blowing.

        If the rock casts a shadow, the sun is shining.

        If the rock does not cast a shadow and is not wet, the sky is cloudy.

        If the rock is difficult to see, it is foggy.

        If the rock is white, it is snowing.

        If the rock is coated with ice, there is a frost.

        If the ice is thick, it's a heavy frost.

        If the rock is bouncing, there is an earthquake.

        If the rock is under water, there is a flood.

        If the rock is warm, it is sunny.

        If the rock is missing, there is a tornado.

        If the rock is wet and swinging violently, there is a hurricane.

        If the rock can be felt but not seen, it is night time.

        If the rock has white splats on it, watch out for birds.

        If the rock is levitating, you're stoned.

        If there are two rocks, you're drunk.

  2. Anonymous Coward
    Anonymous Coward

    I'm happy as long as everyone keeps using different models, I'm happy.

    I know deep down I shouldn't do this, but I usually look at the weather forecast from several sources and decide which I like the best.

    1. David 132 Silver badge

      Re: I'm happy as long as everyone keeps using different models, I'm happy.

      The forecast for this evening/tonight where I am is full of dire warnings of thunderstorms, gusting winds, and hailstones the size of golfballs.

      So I spent a long-overdue hour cleaning my wife’s #%^@! out of the garage to allow me to put my car away (hers has been squeezed into one of the barns) juuuuuust in case.

      Can’t decide if in the event of no hail I will be annoyed or relieved, after going to all that effort!

      But glass half full, I now have that rarest and most precious of things - a garage with a car in it rather than just junk.

    2. LybsterRoy Silver badge

      Re: I'm happy as long as everyone keeps using different models, I'm happy.

      I'm old enough to remember the "quick change over to ITV it might be better there"

      Also I remember blaming the weatherperson for what was comming.

  3. Yorick Hunt Silver badge
    Trollface

    If it ain't broke, don't fix it.

    I still prefer the weather stone, been accurate since its creation;

    https://en.wikipedia.org/wiki/Weather_rock

    1. jake Silver badge

      Re: If it ain't broke, don't fix it.

      That's not forecasting, that's a form of local nowcasting.

    2. Lazlo Woodbine Silver badge

      Re: If it ain't broke, don't fix it.

      Without using a time machine, the weather stone can only inform you of the weather now, or in the very recent past, this is very much not a forecast...

      1. LybsterRoy Silver badge

        Re: If it ain't broke, don't fix it.

        Didn't you know a tripod, a bit of string and a stone IS how you build a time machine. Haven't you seen what the TARDIS looked like before its chameleon field got stuck?

  4. Charlie Clark Silver badge

    Less is more

    Can't read the paper, but this looks like another example that picking the right parameters is key to any statistical exercise, which underpin most of modern AI systems. Starting from relatively simple machine learning, LLMs just added parameters as they came along in the race to be "first". Having now more or less reached the "complexity horizon", other approaches are required.

    1. Anonymous Coward
      Anonymous Coward

      Re: Less is more, 'more or less' !!! :)

      Thank you !!!

      'AI' has run into the 'Complexity Horizon' at full tilt ... and discovered that it is surprisingly 'Hard' in all senses, like a 'gag' from 'Looney Tunes'.

      The resulting 'collision' has resulted in an 'Headache' we are ALL feeling !!!

      Maybe when our 'collective heads' clear we will recover enough sense to recognize that it is all an 'oversold' scam !!!

      [One can hope that the 'bubble' WILL burst and sooner rather than later !!!]

      :)

      1. LionelB Silver badge

        Re: Less is more, 'more or less' !!! :)

        Yerrrs, but doesn't this weather forecasting system represent a move to addressing that Complexity Horizon (or at least pushing it a little further into the future)?

        They appear to be taking on board that, with a bit of domain-specific knowledge and input you can do better than the current default of just throwing a shit-ton of stuff into a ginormous ML.

        Perhaps the real future of ML is not so much tilting at AGI, but effective domain-specific systems.

  5. jake Silver badge

    I use the Golden Gate Bridge.

    I'm only accurate out about four days, but some old-timers get pretty good accuracy out almost a week.

    It works by using the bridge as a kind of barometer. The fog hits the bridge at different heights depending on barometric highs and lows, Their relative movements can be visually tracked, both how far off the water, and how much of the towers can be seen above it. Couple that with observing the wind direction, if one tower is clear and the other is in fog (changing wind directions) and various other factors combine to make for a very useful tool to the observant local. Knowing what the water temp readings readings are at the off-shore buoys are help, too, but some say that's cheating.

    1. Anonymous Coward
      Anonymous Coward

      Re: I use the Golden Gate Bridge.

      I'm surprised that gives accurate forecasts for the UK regions but I trust people on the internet.

      1. Doctor Syntax Silver badge

        Re: I use the Golden Gate Bridge.

        In upland Britain the simple equivalent is "If you can see $LOCAL HILL it's going to rain. If you can't see it it's raining." Maybe those living in boring flat areas have alternatives.

        1. Androgynous Cupboard Silver badge

          Re: I use the Golden Gate Bridge.

          In boring flat areas based on recent data, I believe it's: if your basement is a foot underwater, it's been raining.

          1. Andrew Scott Bronze badge

            Re: I use the Golden Gate Bridge.

            if you're suddenly under water in a slot canyon in arizona it probably rained north of you a few hours ago. Local weather was fine,

        2. DJV Silver badge

          Re: I use the Golden Gate Bridge.

          Yes, I live in a mostly boring flat area (Norfolk) though, as I live on one of the more bumpier parts, I can either get a clear view of the Norwich Southern Bypass as it passes over the river or not depending on the weather (or, in summer, possibly the local trees).

          One of my cats will bury herself underneath the top covers on my bed if it is or is going to be cold. Not exactly forecasting but probably as good as any aardvark that I know of!

          1. John Smith 19 Gold badge
            Coat

            "One of my cats will bury herself underneath the top covers on my bed "

            Is your cat called "The Lord," by any chance?

            Yes, mine's the one with a copy of THGTTG in the side pocket.

        3. LionelB Silver badge

          Re: I use the Golden Gate Bridge.

          I use the volume and timbre of the shunting noises from the railway station about a mile away. I don't know why, but it works quite well, and is weirdly soothing.

  6. Julz

    Why

    Not run the model on the existing weather super computers for more accuracy?

  7. Doctor Syntax Silver badge

    As climate shifts the training data will become increasingly out of date. Basic principles of atmospheric and oceanic physics won't.

    1. Ken Hagan Gold badge

      True, but probably only on timescales much longer than the re-training time for the model.

    2. Charlie Clark Silver badge

      The ERA data is continually being updated and retraining costs, says once a quarter would seem acceptable, especially if you get a system that allows you download trained data sets.

  8. Anonymous Coward
    Anonymous Coward

    We're entering a new era ...

    I've just started with a company that deals with building management systems. Worldwide.

    Thousands of buildings, and hundreds of thousands of data points (outside temperature is a given, humidity is common) that can be pulled together in a second.

    That is a shitload of "free" data to work with.

    I have already suggested throwing it into a learning engine to see what happens

  9. John Smith 19 Gold badge
    Unhappy

    Hmmm. It's *very* impressive sounding

    But, not to put a damper on things how the f**k does it work?

    Now I can see AI simulating how a human weather forecaster does it, but I'm struggling with

    multi-layer neural network --> models human forecasting process ---> new forecast.

    1. Ken Hagan Gold badge

      Re: Hmmm. It's *very* impressive sounding

      I doubt anyone knows, or expects to know. A neural net isn't an algorithm. Although the different stages may often be inspired by the kinds of categorisation that humans apply when doing the same task "by hand", there's no hard and fast rule requiring this. You could just start with a layer containing all the raw measurements, end with a layer delivering the time-series of the forecasts for different places, and stick a few layers of "mix it all up" in between and see how you get on.

      If it works, it's useful. Eventually, one might study the best of the working models and see if the layers are doing something specific that no-one had thought to try before. It could give you clues that eventually lead to genuine scientific progress. In the meantime, not needing a super-computer to create a forecast might even save you money. (Perhaps not, mind, since you probably do still need something quite chunky for the training.)

      1. LionelB Silver badge

        Re: Hmmm. It's *very* impressive sounding

        Indeed. There is in fact already a considerable academic push, though, (including at my own research centre) into developing techniques to "peer into the black box" and gain insight into how opaque ML systems (yes, they are algorithms!) achieve the functionality that they do.

    2. Filippo Silver badge

      Re: Hmmm. It's *very* impressive sounding

      A neural network is essentially a system that can encode patterns in data during the training phase, and then recognize them during the inference phase. Meteorological data has patterns, and those patterns correlate with future weather.

      It works similarly to the old folk who know nothing of atmospheric physics, but have observed the weather in their town every day for fifty years. They can probably make a decent forecast on just the data they get from their senses, although they won't be able to explain it much better than "because it's always like that". That literally means "because the current data matches a pattern I recognize".

      The NN works in a similar fashion, except that it has observed a lot more data on a lot more weather in a lot more locations.

      1. Doctor Syntax Silver badge

        Re: Hmmm. It's *very* impressive sounding

        "It works similarly to the old folk who know nothing of atmospheric physics"

        The old Red sky at nigh is a very generalised version of this. It works because the weather in the UK usually arrives from the west. What it doesn't say is that it won't work when we're in a run of easterlies. If this is what ML accomplishes it will have the same limitations. As someone said up-thread predicting tomorrow will be like today has a pretty good success rate without any form of computation.

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Other stories you might like