When Adaptive Rewards Hurt: Causal Probing and the Switching-Stability Dilemma in LLM-Guided LEO Satellite Scheduling

1. 1Identify the problem of adaptive rewards in deep reinforcement learning
2. 2Understand the switching-stability dilemma and its impact on performance
3. 3Evaluate the trade-offs between static and dynamic reward weights
4. 4Consider the requirements of quasistationary reward signals for value function estimation

Machine learning researchers and engineers working on reinforcement learning and satellite scheduling can benefit from understanding the limitations of adaptive rewards, as it can inform their design choices and improve the performance of their models

💡 Near-constant reward weights can outperform dynamic weights due to the need for quasistationary reward signals in PPO

💡 Adaptive rewards can hurt performance in LEO satellite scheduling due to switching-stability dilemma