Artificial Intelligence: applying ‚Deep Reinforcement Learning‘ for sustainable development

20/12/2019 - For the first time, a specific way of machine learning has been used to find novel pathways for sustainable development. So far, the so-called 'Deep Reinforcement Learning' has mostly been used to make computers excel in certain games, such as AlphaGo, or navigate robots through rough terrain. Now, scientists from the Potsdam Institute for Climate Impact Research developed a mathematical framework combining recently developed machine learning techniques with more classical analysis of trajectories in computer simulations of the global climate system and the global economy. The results, published in the interdisciplinary journal on nonlinear phenomena 'Chaos', are promising.
Artificial Intelligence: applying ‚Deep Reinforcement Learning‘ for sustainable development
Dynamics of a stylized World-Earth system. Cutout from Figure 3 of Strnad et al, 2019

The scientists are stressing that the model of our world they used for testing their way of ‘Deep Reinforcement Learning’ is deliberately simplified. Hence the pathways for sustainable development identified by the artificial intelligence in the model, including for instance a specific mix of CO2 emissions taxation and subsidies for renewable energies, cannot directly be used in the real world. Yet the results show that the scientists’ application of machine learning does indeed find pathways that are innovative, compared to what classical analysis yielded. The study is hence of substantial methodological value. The authors call upon other scientists to apply their tool to further models in order to learn more about the potential of machine learning for sustainable development.

Article: Felix M. Strnad, Wolfram Barfuss, Jonathan Donges, Jobst Heitzig (2019): Deep reinforcement learning in World-Earth system models to discover sustainable management strategies. Chaos (Vol.29, Issue 12) ]DOI: 10.1063/1.5124673]

Weblink to the article: