How to predict extreme catastrophes when their low frequency prevents feeding algorithms? American researchers believe they found the solution by combining advanced machine learning techniques and sequential sampling.
The cost of natural disasters in France will reach almost 10 billion euros in 2022, France Assureurs announced at the end of January. A record since 1999 with the “intensification of extreme climate phenomena”.
At the same time, the British NGO Christian Aid published a report on a global level in which it estimated that the ten largest natural disasters in 2022 cost more than 158 billion euros.
In order to limit the effects of these extreme events (very strong earthquakes, pandemics, rogue waves, etc.), scientists try to predict their formation and their response. Easier said than done ! These extreme phenomena are by definition rare, even if France Assureurs notes an “increase in frequency”.
Statistically, there isn’t enough data about them to use predictive models to accurately predict when they will occur. But a group of scientists from Brown University and the Massachusetts Institute of Technology may have found an effective method.
In a study published in Nature Computational Science in late 2022, the researchers explain how they used statistical algorithms that require less data to make accurate predictions, combined with a powerful machine learning technique developed at Brown University.
This combination allowed them to predict scenarios, probabilities, and even timelines of rare events despite the lack of historical data. To simplify things, they sidestepped the problem of the lack of large amounts of data to focus on the quality rather than the quantity of the information.
So they asked themselves, “What is the best possible data we can use to minimize the number of data points needed?” »
Anticipating rogue waves
The researchers found the answer in a sequential sampling technique called active learning. Not only are these types of statistical algorithms capable of analyzing the data given them, but more importantly, they can learn from that information to identify new relevant data points that are just as important.
For several years, active learning has been used in natural language processing, where unlabeled data can be plentiful but annotation is slow and expensive.
Scientists from Brown University and MIT therefore used this active learning technique to feed their machine learning model and proactively look for clues that point to rare events.
Introduced in 2019 by these Brown researchers, DeepOnet (Deep Operator Networks) is a type of artificial neural network that uses nodes connected to each other in successive layers, roughly mimicking the connections made by neurons in the human brain.
According to this team, DeepOnet can be trained to find the parameters, or precursors, that lead to the catastrophic event being analyzed, even when there are few data points.
For example, regarding rogue waves, researchers have found that they can determine and quantify when they form by looking at the likely conditions of waves interacting in nonlinear ways over time, resulting in waves that are sometimes three times larger than their original size.
The team is currently working with environmental scientists to use their method to predict weather events such as hurricanes.
In documentary sources