Can We Design an AlphaGo to Forecast Floods?

This year’s Climate and Society class is out in the field (or lab or office) completing a summer internship or thesis. They’ll be documenting their experiences one blog post at a time. Read on to see what they’re up to.

Jiayu Xu, C+S ’17

Machine learning and artificial intelligence are some of the hottest topics in the past few decades. They have caught wide attention because their applications as diverse as data mining, self-driving cars, fingerprint identification, financial services and other fields.

The public become more familiar with machine learning after the computer program AlphaGo defeated the legendary Go champion Lee Sedol during a match in 2016. The intelligence of machines have shocked human, especially their outstanding abilities of self-learning, dealing with large quantities of data and predict the future. A logical next step is for researchers to use machine learning to forecast the weather. Is it possible? Can we design an “AlphaGo” to better predict floods or hurricanes? Let’s see…

Go play miracle: AlphaGo

Before we learn about AlphaGo, here’s a quick look at the definition of artificial intelligence and machine learning. Concisely, AI is the ability of machines and computer systems to simulate human intelligence. Machine learning is a branch of artificial intelligence, and its main idea is that machines should have the ability to learn from datasets and adapt through experience by themselves rather than relying on designed programming codes. Its main purpose is to predict the future through learning data in the past, formulating hypotheses, refining and testing hypotheses again and again.

Invented in ancient China, Go is a strategy board game with the aim of surrounding a larger area on the board than your opponent, which is more complex than chess. In May 2016, AlphaGo — a computer program developed by Google DeepMind — defeated 18-time world champion Lee Sedol during by a score of 4-1. Even Sedol was impressed by the AlphaGo’s skills.

Confirming that a computer program had the ability to surpass human intelligence in playing Go — one of the most challenging mind sports in the world — was an extraordinary leap for the AI field.

Machine learning and neural networks are crucial for making the computer program so powerful because they train it to mimic human play. In other words, AlphaGo can teach and play millions of games of Go with itself, and analyze the probability of winning for each move much faster and more accurately than the human brain does, which makes it possible to predict the best move during a Go match.

Board Game (Source: Wikipedia/Goban1)

Extend the miracle to meteorology?

The application of machine learning is not limited to Go. Its value and potential has been widely recognized by many industries that deal with large amounts of data. For example, financial organizations use machine learning to identify investment opportunities and the right time to trade. By being trained with millions of historical data points and experience, machines are able to formulate hypotheses and predict the prospective data trends.

Beyond finance, this approach could be perfect to improve meteorological forecasts.

During my undergraduate years studying atmospheric sciences, I was taught that most weather prediction models are based on a series of physical equations that are used to explain what’s going on in the atmosphere and oceans.

In fact, the atmosphere is much more complicated in real life, making difficult to build a perfect forecast model. Machine learning can depend on automatic learning to become more accurate in predicting outcomes without being limited by equations, though. That means it could be reasonable to design a powerful tool like AlphaGo to improve the weather forecast. In fact, plenty of academic institutions, commercial companies and nonprofit organizations have been working on it, including Stanford University, IBM, and the Red Cross.

Machine Learning for Flood Forecasts

As a junior researcher intern for Red Cross Red Crescent Climate Centre this summer, I’m fortunate to be taking a deep dive into machine learning. One of my main tasks is to assist in improving the development of the FUNES, a prototype machine learning tool. Our team plans to test its ability to anticipate patterns of flooding events in Togo based on available data.

In addition to testing the predictive performance of FUNES by adjusting the gap and threshold of variables such as rainfall and flood flow rates, I have collected soil moisture datasets for the FUNES and tested whether they have reasonable correlations with other data, since the saturated soil can help trigger a flood.

The correlation between soil moisture and flood flow in FUNES (Source: Jiayu Xu)

FUNES is a newborn machine learning tool so we are finding as many datasets as possible to train it and enhancing its forecast capability. Every time FUNES has been updated to a new version based on our efforts and feedback, I feel thrilled by its progress and potential. Our supervisor told us that if FUNES works out, it can eventually be used by Red Cross and partners to forecast extreme events. I really hope FUNES can be more powerful and intelligent in the future, and maybe become the AlphaGo for forecasting floods.

Machine learning is likely to change the world in the near future and amaze the public over and over again. Meteorology is only one of fields where it has the potential shine brilliantly.