Overview
Mobile base stations (MBSs), comprising both aerial and ground-based platforms, offer flexible wireless coverage through adaptive positioning. Realizing this flexibility requires joint optimization of physical trajectories and digital beamforming. To resolve the coupled operational challenges and high interaction costs of standard model-free reinforcement learning, researchers formulated the task as a constrained Markov decision process (CMDP) and introduced a SafeDreamer-based world model framework. The method learns latent system dynamics from environment experience and optimizes joint control policies across imagined trajectories while enforcing mobility, restricted-zone avoidance, and transmit-power constraints.
Research Context
The joint control of trajectory and digital beamforming in MBS platforms presents closely coupled physical and electromagnetic dynamics:
- Platform movement continuously changes wireless channel conditions and spatial interference patterns.
- Digital beamforming decisions directly determine the resulting signal and interference power levels.
- Current positional movements dictate future spatial locations, communication performance, and adherence to mobility and safety constraints.
Model-free reinforcement learning algorithms typically demand extensive real-world or simulated environment interactions to navigate these continuous feedback loops. Furthermore, operating mobile transmitters requires strict adherence to physical mobility boundaries, restricted-zone avoidance, and maximum transmit-power budgets, complicating policy optimization in unconstrained learning architectures.
Approach
The problem is cast as a CMDP aimed at maximizing the time-averaged sum signal-to-interference-plus-noise ratio (SINR). The formulation explicitly incorporates constraints covering:
- Platform mobility limits.
- Restricted-zone avoidance.
- Maximum transmit-power constraints.
To reduce interaction overhead and support long-term planning, the authors implemented a model-based framework rooted in SafeDreamer. The architecture learns latent representations of the underlying system dynamics from collected experience. Rather than interacting continuously with the raw environment, the joint trajectory and digital beamforming policy is optimized across internally imagined trajectories.
Operational constraints are integrated directly into the policy learning pipeline through two primary mechanisms: feasibility-preserving action mappings to handle structural boundaries, and dedicated safety cost signals that penalize constraint violations during policy updates.
Findings
Simulation evaluations demonstrated concrete performance and efficiency gains over standard reinforcement learning methods:
- The proposed world model framework improved the time-averaged sum SINR by an average of 7% compared with a proximal policy baseline.
- The framework achieved an average 9% improvement in time-averaged sum SINR compared with a deep Q-network baseline.
- The system demonstrated higher sample efficiency during training by leveraging dynamics learned in latent space.
- Constraint satisfaction was more reliably achieved relative to the comparative baselines through the combination of safety cost signals and feasibility-preserving action mappings.
Why This Matters
Joint trajectory and beamforming optimization allows aerial and ground mobile base stations to adaptively tailor coverage to dynamic environments. By replacing direct trial-and-error interactions with latent trajectory imagination, the safe world model framework provides a mechanism to improve network SINR while maintaining adherence to transmit-power limits, physical mobility boundaries, and restricted exclusion zones.