LAMPS: A Learning-based Mobility Planning via Posterior \\ State Inference using Gaussian Cox Process Models

Egemen Erbayat, Yongsheng Mei, Gina Adam, Suresh Subramaniam, Sean Coffey, Nathaniel D. Bastian, Tian Lan

Keywords: Mobility planning, kernel density estimation, reinforcement learning, multi-agent system

Abstract: Learning-based mobility planning has proven effective in optimizing performance metrics like latency, throughput, and cost in applications such as path planning and network security. However, real-world networks often face partial or dynamic observability, limiting the applicability of existing robust optimization approaches, which can be conservative, inefficient, or require extensive retraining under changing conditions. This paper introduces LAMPS, a new learning-based mobility planning framework that leverages Gaussian Cox processes to estimate spatiotemporal network states and their uncertainty, enabling robust decision-making under varying observability without retraining. These posterior estimates are integrated into a utility-based planning algorithm that adapts policies trained under full observability to diverse conditions, optimizing average performance or ensuring robustness in near-worst-case scenarios. We analyze the LAMPS framework in a real-world situation involving UAV mobility and wireless resource management, demonstrating the framework’s scalability, adaptability, and efficiency in dynamic network environments. Our evaluation results demonstrate the remarkable adaptability and robustness of the LAMPS. It consistently outperforms other methods under different observability conditions and setups, proving its effectiveness in dynamic environments without requiring retraining.

Submission Number: 11