For the longest time, the basic approach to developing an AI was for the humans to teach the machine everything they could, then have the software take it from there. That approach worked. It's how IBM's Deep Blue defeated world chess champion Garry Kasparov in a six-game match in 1997 and how Google's DeepMind's Alpha Go defeated arguably the world's top Go player in five games in 2016.
Then the scientists had a different idea: What if they let the AI learn entirely on its own, without regard for any human preconceptions, after just being given the rules of a game? That worked even better. By playing millions of games against itself, what DeepMind called Alpha Go Zero learned Go so well in three days that it defeated Alpha Go in 100 straight games.
DeepMind then went the next step and developed an AI that hadn't even been taught the rules of Go. It trounced Alpha Go Zero.
DeepMind is taking that sort of approach with hurricane forecasting. Rather than use the traditional approach — feeding massive amounts of data to supercomputers loaded with physics equations that spend hours and hours calculating forecasts for storms — DeepMind left out the physics equations piece, as well as all other guidance. Basically, DeepMind says: Here is all the data we have on hurricanes. You figure out what it means for future storms.
The approach has shown promise with earlier storms, and DeepMind's AI just nailed the forecast for Hurricane Erin, outperforming both the official, supercomputer-based forecast and other commonly used models.
Let's have a look at how far the AIs have come, so very fast, as well as where they can go from here.
The promises of the deep learning approach first showed up on my radar not quite two years ago. In September 2023, I wrote a commentary lauding what advancements in supercomputing and satellite imagery were doing for forecasting. Just a month later, I found myself writing about AI models that, according to the Washington Post, had shown during that hurricane season that they "portend a potential sea change in how weather forecasts are made."
Now, Ars Technica reports that Google's AI outperformed the official forecast and numerous other of the best physics-based models on both intensity and the storm track, even after the other models were corrected for known biases.
The article notes that the outperformance occurred with predictions reaching out to as much as three days ahead, while the most important forecasts are those three to five days ahead, because that's when many key decisions about evacuations and other preparations are being made.
"Nevertheless," Ars Technica says, "the key takeaway here is that AI weather modeling is continuing to make important strides. As forecasters look to make predictions about high-impact events like hurricanes, AI weather models are quickly becoming a very important tool in our arsenal.
"This doesn't mean Google's model will be the best for every storm. In fact, that is very unlikely. But we certainly will be giving it more weight in the future.
"Moreover, these are very new tools. Google's Weather Lab, along with a handful of other AI weather models, has already shown equivalent skill to the best physics-based models in a short time. If these models improve further, they may very well become the gold standard for certain types of weather prediction."
Let's hope that the AIs continue their remarkable progress and, if so, that the public comes to trust them. A lot of damage and injury could be avoided.
In the meantime, fingers crossed that this year's hurricane season stays relatively quiet.
Cheers,
Paul