Unverified50% confidenceFactExact time
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
1
Sources
50%
Confidence
Long-term
Relevance
8/5/2026
First Seen
Sources
Related Claims
Unverified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73% similarUnverified传统自回归推理中，模型每一步只能基于已有上下文预测下一个token，是串行依赖关系67% similarUnverifiedOn divergent tasks requiring multi-step reasoning, long-context understanding, or open-ended architectural decisions, the advantages of newer models may be significant.62% 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.61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/690913API
curl https://kongchang.com/api/v1/knowledge/claims/690913MCP
get_claim(id=690913)