Unverified50% confidenceFactExact time
Policy gradient methods directly parameterize the policy π(a|s) and maximize expected return through gradient ascent, naturally supporting continuous action spaces and stochastic policies, with representative algorithms including REINFORCE and PPO.
1
Sources
50%
Confidence
Long-term
Relevance
10/26/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedFor continuous action spaces, PPO is typically paired with a Gaussian policy output, where the network simultaneously predicts the mean and standard deviation of actions.75% similarUnverified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74% similarUnverifiedThe occupancy measure d^π(s,a) characterizes the expected number of visits or the normalized distribution of a policy π visiting state-action pairs (s,a) in an MDP, and in episodic MDPs it simultaneously encodes the long-term behavior of the policy.74% similarUnverified策略梯度定理表明期望回报对策略参数的梯度可表示为 ∇J(θ) = E[∇log π_θ(a|s) · A(s,a)]，其中A是优势函数72% similarUnverified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72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1100160API
curl https://kongchang.com/api/v1/knowledge/claims/1100160MCP
get_claim(id=1100160)