Unverified50% confidenceFactExact time
POMDP在MDP基础上增加了观测函数O(o|s,a)，智能体只能获得状态的部分信息，需要维护对真实状态的置信度
1
Sources
50%
Confidence
Long-term
Relevance
8/7/2026
First Seen
Sources
M.A.R.A：用强化学习从零训练会战斗的AI坦克
Related Claims
UnverifiedPOMDP（部分可观测马尔可夫决策过程）是MDP（马尔可夫决策过程）的推广形式，在标准MDP中智能体能完全观测环境状态，而POMDP承认智能体无法直接获知真实状态，只能通过带噪声的观测间接推断74% similarUnverifiedPOMDP introduces the belief state, i.e., a probability distribution defined over all possible hidden states, based on which the agent makes decisions rather than from a single deterministic state.67% similarUnverified贝叶斯滤波是POMDP中维护信念状态的核心机制，通过状态转移模型预测先验概率、用观测模型计算似然、归一化得到后验信念65% similarUnverified当序贯决策问题同时具备隐藏状态、带噪声的观测和序贯决策三个要素时，适合用POMDP建模62% similarUnverified自注意力机制的QK^T计算对于序列长度n、隐藏维度d的模型需要O(n²d)次乘加运算61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/699557API
curl https://kongchang.com/api/v1/knowledge/claims/699557MCP
get_claim(id=699557)