Unverified50% confidenceFactExact time
策略梯度的核心思想源自 REINFORCE 算法（Williams, 1992），通过对轨迹回报的蒙特卡洛估计来计算策略参数的梯度
1
Sources
50%
Confidence
Long-term
Relevance
8/7/2026
First Seen
Sources
Related Claims
Unverified策略梯度方法本质上是on-policy的，因为梯度估计中的期望是关于当前策略分布的73% similarUnverified贝尔曼方程将长期累积奖励的优化问题分解为可递归求解的子问题，奠定了Q-learning、策略梯度等经典算法的理论根基67% similarUnverifiedThe backpropagation algorithm was systematically proposed by Rumelhart, Hinton, and Williams in 1986, and its core relies on the chain rule to propagate the gradient of the loss function layer by layer to each parameter.66% similarUnverifiedBayesian optimization constructs a probabilistic surrogate model of the objective function (typically a Gaussian process) and uses an acquisition function (such as Expected Improvement or Upper Confidence Bound) to determine the next sampling point.64% similarUnverifiedBayesian optimization uses Gaussian processes to model the objective function for hyperparameter tuning.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/705622API
curl https://kongchang.com/api/v1/knowledge/claims/705622MCP
get_claim(id=705622)