Item

Abstract

Deep reinforcement learning has empowered recent advances in games like chess or in language modelling with chatGPT, but can also control nuclear fusion in a Tokamak reactor. In the often-used actor-critic framework, a neural network is trained to control while another neural network evaluates the actions of the first network. In this talk, we cast model error correction into a remarkably similar framework to learn from temporally sparse observations. A first neural network corrects model errors, while a second, simultaneously trained, estimates the future costs if the model error correction were applied. This allows us to circumvent the need for the model adjoint or any linear approximation for learning in a gradient-based optimization framework.We test this novel framework on low-order Lorenz and sea-ice models. Trained on already existing trajectories, the actor-critic framework can not only correct persisting model errors, but significantly surpasses linear and ensemble methods. Furthermore, using this framework enables us to learn unknown processes in sea-ice models from temporally sparse observations. Therefore, we see training of model error corrections with such an actor-critic framework as one of the most promising steps for geoscientific hybrid modelling.