I am very seldom interested in applications. I am more interested in the elegance of a problem. Is it a good problem, an interesting problem? - Claude Shannon

Climate change is a tractable problem. For one thing, we have the internet. Maybe good things can go viral. Humans are numerous and the cause of climate change. Thus if humans want to stop climate change they can.

Machine learning is one useful tool. Many people study machine learning. Nonetheless, machine learning is not so much a field itself so much as a tool used in other fields. Thus a multitude of fields contribute and collaborate towards machine learning problems.

ML problems diagram

Who do we want to convince that machine learning can be used for something other than video games, smart homes for the elderly, and superintelligence? Researchers, engineers, entrepreneurs, investors, big companies, and governments. There are machine learning-climate change problems that can be targeted at each audience.

There are a variety of different ML problems. Some are more short-term than others. We can categorize them into three categories. ‘High Leverage’, meaning straight up good problems. ‘Long-term’, meaning problems with main impact after 2040. ‘High Risk’, meaning high risk and uncertainty of effectiveness.

You should collaborate with domain experts when working on an ML problem in that domain.

Theoretically solving climate change will result in cool theories. We will need policies to stop climate change in real life.

AI has been described as the new electricity, which is pretty stupid because without electricity it would be cold and dark. However, AI can improve electricity, which would help with climate change because less electricity means less coal. We want to transfer from traditional energy sources such as coal and fossils to alternative “low carbon” sources such as solar, wind, hydro, and nuclear. We still want to reduce current emissions. This should be applied outside just America. ML can help with this transition

There are two types of alternative energies, variable and controllable.

Current electric grids cannot handle variable energies without wasting extra CO2, since they cannot store power for later. ML can both help reduce the CO2 currently wasted and transition to alternative energies, by improving technology and intelligently accommodating the variability of both supply and demand. Predicting supply and demand is one way to do so.

Currently, when balancing power systems, operators use a slow and complex processes governed by NP-hard optimization problems. Machine learning techniques can help solve these optimization problems, or bypass them with direct control learned by subfields like reinforcement learning.

Read the rest here on Arxiv.