A Unified Framework for Dynamic Reward Shaping in Reinforcement Learning

Dynamic reward shaping architectures must account for evolving value estimates and feedback loops during training to avoid policy divergence. This framework addresses sparse and delayed signals by integrating adaptive auxiliary rewards that adjust alongside the learner's improved predictive models and changing information density.