Unverified50% confidenceFactExact time
基于势函数的奖励塑形（Ng 等人，1999）经理论证明不会改变最优策略，其形式为 F = γΦ(s') - Φ(s)
1
Sources
50%
Confidence
Long-term
Relevance
7/11/2026
First Seen
Sources
Reward Shaping and RL Debugging in Practice: A Systematic Guide from Oscillation to Convergence
redditr/reinforcementlearning7/9/2026
Related Claims
Unverified1999年Ng、Harada与Russell在ICML证明，只有基于势函数的奖励塑形（F(s,s')=γΦ(s')-Φ(s)）才能保证不改变最优策略79% similarUnverifiedHardy、Benedicks、Amrein-Berthier等人证明了函数及其傅里叶变换的支撑集不能同时具有有限测度（除非函数为零）63% similarUnverifiedReward shaping techniques were theoretically analyzed by Andrew Ng and others in a 1999 paper, which proved how reward shaping can be safely performed without altering the optimal policy.62% similarUnverified拉格朗日对偶通过引入拉格朗日乘子将约束吸收进目标函数，再利用 KKT 条件转化为对偶问题，引出支持向量和核技巧的理论基础61% similarUnverifiedHardy在1933年证明：如果函数f(x)和其傅里叶变换都以高斯函数e^{-πx²}速率衰减，则f必须是高斯函数的常数倍；衰减更快则f≡061% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/474923API
curl https://kongchang.com/api/v1/knowledge/claims/474923MCP
get_claim(id=474923)