When it comes to catastrophic events (think of earthquakes, epidemics, or “deceptive waves” that can destroy coastal structures), computational models face almost uncontrollable challenges: It is rare that there is not enough data on them to use predictive models to accurately predict when they will occur next.
But researchers from Brown University and the Massachusetts Institute of Technology say that is not the case.
In a new study at Natural Computational ScienceScientists describe how they combine statistical algorithms – which require less data to make accurate and efficient predictions – with powerful machine learning techniques developed at Brown and trained to Predict the probability status and sometimes even the timing of rare events, even if there is no historical record on them.
In doing so, the research team found that the new framework could provide a way to avoid the need for large amounts of data, which is traditional for these types of calculations, instead of the huge challenge of predicting rare events. As a matter of quality. Excess quantity.
“You have to know that these are stochastic events,” said George Karniadakis, a professor of applied mathematics and engineering at Brown and author of the study. “Outbreaks of epidemics such as COVID-19, environmental disasters in the Gulf of Mexico, earthquakes, wildfires in California, and 30-meter waves that sank ships – these are rare events because But rarely do we. There is a lot of historical data. We do not have enough examples from the past to predict them further in the future. The question we address in the paper is: What? What is the best possible data we can use to reduce the number of data points we need?
Researchers found the answer in a sequential modeling technique called active learning. These types of statistical algorithms can not only analyze data entry into them, but more importantly, they can learn information to label new relevant data points that are equal or more important to the results. Which is being calculated. At the most basic level, they allow to do more with less.
That is important for the machine learning model that the researchers used in the study. Called DeepOnet, this model is a type of artificial neural network that uses interconnected nodes in successive layers that mimic the interactions made by neurons in the human brain. DeepOnet is known as a deep nerve operator. It is more advanced and powerful than a normal artificial neural network because it is actually two neural networks processing data in two parallel networks. This allows it to analyze large data sets and scenarios at split speeds to release an equally large set of probabilities when it knows what it is looking for.
The drawback with this powerful tool, especially as it relates to rare events, is that deep nervous system operators need a lot of data to be trained to perform calculations effectively and accurately.
In the paper, the research team shows that combined with active learning techniques, the DeepOnet model can be trained on parameters or pre-determined factors to look for that lead to a catastrophic event that someone is analyzing, even without a lot of data points. By.
“The impetus is not to take every possible data and put it into the system, but to actively look for events that will reveal rare events,” Karniadakis said. “We may not have many examples of real events, but we can have those causes. “Through mathematics, we identify them, which together with real events will help us to train operators who are hungry for this data.”
In the paper, researchers apply methods to determine the parameters and range of different probabilities for dangerous surges during an epidemic, finding and predicting deceptive waves and estimating when a ship will split in half due to stress. For example, with deceptive waves – waves that are twice the size of the surrounding waves – researchers have found that they can detect and quantify when deceptive waves occur by looking at possible wave conditions. There are those that do not interact with time, leading to waves sometimes three times. Their original size.
Researchers have found that their new approach is more effective than traditional modeling efforts, and they believe it provides a framework that can effectively detect and predict all types of rare events.
In this paper, the research team describes how scientists should design future experiments so that they can reduce costs and increase the accuracy of predictions. Karniadakis, for example, is already working with environmental scientists to use novel methods to predict weather events such as hurricanes.
The study was led by Ethan Pickering and Themistoklis Sapsis from MIT. DeepOnet was introduced in 2019 by Karniadakis and other Brown researchers. They are currently seeking patents for the technology. The study was supported by funding from the Defense Advanced Research Projects Agency, the Air Force Research Laboratory, and the Navy Research Office.
InsurTech Data Boost
Interested in I?
Receive automatic notifications for this topic.