Verified65% confidenceFactExact time
Research by Ng et al. in 1999 proved that only potential-based reward shaping guarantees that the optimal policy remains unchanged.
3
Sources
65%
Confidence
Long-term
Relevance
7/19/2026
First Seen
Sources
温室控制器大比拼：强化学习为何败给规则基线？
Related Claims
UnverifiedThe 1999 work by Ng et al. pointed out that inappropriate reward shaping can change the optimal policy, causing the agent to pursue auxiliary rewards instead of the true objective (reward hacking).76% similarUnverifiedThe study recommends that rather than pursuing a single perfect reward, one should systematically anticipate and test shortcuts the model might exploit, and incorporate adversarial validation into the design phase.72% similarUnverifiedPotential-based reward shaping (PBRS) theoretically guarantees that shaping does not alter the functional form of the optimal policy, but does not cover the issue of magnitude.70% similarUnverifiedReward shaping needs to be designed carefully to prevent the agent from exploiting reward loopholes and producing unintended behavior.69% similarUnverifiedThe study concludes that preference conditioning alone cannot guarantee controllable multi-objective behavior, and that building robust fair RL-MIL systems requires the introduction of explicit gradient balancing, objective separation, and stability analysis mechanisms.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/561528API
curl https://kongchang.com/api/v1/knowledge/claims/561528MCP
get_claim(id=561528)