Pre-train with Path Normalization Intrinsic Reward for Multi-Cell Selection in Mobile Communication

Reinforcement learning (RL) can make dynamic decisions to maximize long-term system utility and has shown promising performance in network communication tasks. However, task-specific RL agents usually require extensive interactions and may adapt inefficiently when downstream reward functions or environment dynamics change, limiting their reuse when downstream tasks change. To address this issue, we propose PPNIR, a world-model pre-training framework for improving downstream adaptation. During reward-free pre-training, PPNIR uses intrinsic rewards based on predictive path differences to encourage exploration of environment dynamics. During fine-tuning, the pre-trained world model is transferred to downstream tasks to improve data efficiency. We further design an Assessment Replay Buffer (ARB) to select a reliable world model according to prediction quality, and introduce a reliability-aware mechanism to dynamically adjust the imagined sequence horizon during policy training. We validate PPNIR in classic image-based RL environments and a communication-oriented CoMP cell-selection environment. The results demonstrate improved data efficiency across downstream tasks with varying rewards and dynamics.