Recasting Oceanic Eddy Trajectory as a Markov Decision Process

Eddy trajectory prediction is crucial for understanding ocean dynamics and has practical implications in various fields. While existing sequential modeling approaches potentially introduce redundancy, this study reformulates oceanic eddy trajectory as a Markov Decision Process (MDP), assuming that the next position of an eddy can be predicted based solely on its current state, rather than on extended historical sequences. Under this framework, a simple Multi-Layer Perceptron (MLP) model, optimized by minimizing the difference between predicted and actual eddy propagation velocities, achieves prediction accuracy comparable to that of more complex sequential modeling approaches. This not only validates the Markovian assumption, but also provides a computationally efficient alternative for eddy forecasting. Furthermore, we identify a critical issue in existing prediction evaluations on employing retrospective state variables as model inputs for intermediate forecast days. The problem is addressed by reusing the initial-day state variables throughout the entire forecast period, thereby establishing a more reliable benchmark for future research.