Reward-Shaping Engineering

When to Use This Skill

Invoke this skill when you encounter:

• Reward Design: "How do I design the reward function for my task?"
• Slow Learning: "Training is extremely slow with sparse rewards" or "Dense rewards cause weird behavior"
• Reward Hacking: "My agent learned a trick that works in training but fails on test", "Agent oscillating instead of balancing"
• Potential-Based Shaping: "How to add shaping without breaking the optimal policy?"
• Distance-Based Rewards: "How to reward progress toward goal without hacking?"
• Inverse RL: "I have expert demonstrations, can I learn reward from them?"
• Auxiliary Rewards: "Should I add helper rewards like action smoothness?"
• Reward Scale Issues: "Training feels different when rewards change magnitude"
• Sparse vs Dense: "When should I use sparse vs dense rewards?"
• Reward Validation: "How do I verify my reward function is correct?"
• Credit Assignment: "How to help agent understand which actions led to success?"
• Normalization: "Should I clip or normalize rewards?"