A Decision Support Tool for Optimal Control of Planet Temperature Using Reinforcement Learning

Climate change and the rise of the planet temperature can have significant impacts on the natural environment and the human society. Temperature regulation recently becomes an important global issue in order to ensure a sustainable climate environment. A major challenge in controlling the atmospheric temperature is that the process is complex and only approximately known.
This paper proposes a decision making support tool to regulate the planet temperature in the presence of seasonal effects, green house gas fluctuations and measurement noises. Reinforcement-learning algorithm was used to construct the decision support tool. The algorithm learns how to optimally affect the temperature of the planet based on its interaction with the environment through changing carbon dioxide levels. No knowledge of the climate system is required. The decision support tool includes an actor and a critic. By using data from past temperature measurements and carbon dioxide levels, the critic updates the action value function, which is then used by the actor to calculate an optimal carbon dioxide level. A climate model with simulated measurement noises and uncertainties was used to evaluate the performance of the decision support tool. Two scenarios were constructed in the simulations. In the first scenario, the methane level of the planet was increased and fluctuated. Due to the effects of this fluctuation, the atmospheric temperature would be significantly changed and might lead to unstable climate environment on the planet, if no control effort was made. In the second scenario, the planet temperature is assumed to have already been high. An attempt to lower the global temperature was made by following a carbon dioxide profile proposed by the decision support tool.
Simulation results show that the decision support tool optimally adjusted and regulated the planet temperature by minimizing both the error (the difference between the desired and the actual measured temperature) and the change in carbon dioxide level. Even with large rises and falls in global atmospheric methane levels, changes in yearly average temperature of the planet were kept within two degrees Celsius in the closed loop with the decision support tool, compared to six degrees Celsius in the open loop (without using any control algorithm). In the scenario where the planet temperature is initially high, following the decision support tool leads to a smoothly reduction of the temperature to the desired level in approximately six years with the least possible required changes in the carbon dioxide level.