Abstract
Conservation laws are considered to be fundamental laws of nature. It hasbroad applications in many fields, including physics, chemistry, biology,geology, and engineering. Solving the differential equations associated withconservation laws is a major branch in computational mathematics. The recentsuccess of machine learning, especially deep learning in areas such as computervision and natural language processing, has attracted a lot of attention fromthe community of computational mathematics and inspired many intriguing worksin combining machine learning with traditional methods. In this paper, we arethe first to view numerical PDE solvers as an MDP and to use (deep) RL to learnnew solvers. As proof of concept, we focus on 1-dimensional scalar conservationlaws. We deploy the machinery of deep reinforcement learning to train a policynetwork that can decide on how the numerical solutions should be approximatedin a sequential and spatial-temporal adaptive manner. We will show that theproblem of solving conservation laws can be naturally viewed as a sequentialdecision-making process, and the numerical schemes learned in such a way caneasily enforce long-term accuracy. Furthermore, the learned policy network iscarefully designed to determine a good local discrete approximation based onthe current state of the solution, which essentially makes the proposed methoda meta-learning approach. In other words, the proposed method is capable oflearning how to discretize for a given situation mimicking human experts.Finally, we will provide details on how the policy network is trained, how wellit performs compared with some state-of-the-art numerical solvers such as WENOschemes, and supervised learning based approach L3D and PINN, and how well itgeneralizes.