Unverified50% confidenceFactExact time
DQN是off-policy算法，可以重复利用过去收集的经验进行学习
1
Sources
50%
Confidence
Long-term
Relevance
8/20/2026
First Seen
Sources
Related Claims
Unverified经验回放（Experience Replay）由DeepMind在2013年DQN论文中引入，将状态-动作-奖励-下一状态四元组存入回放缓冲区，训练时随机采样以打破时序相关性66% similarUnverified潜在缓存技术将LLM推理中的KV Cache思想迁移至训练阶段，对轨迹历史帧预先提取潜在表示并缓存62% similarUnverified在离线强化学习算法（如 Decision Transformer、IQL）中，episode 边界信息对于正确计算回报（return-to-go）至关重要62% similarUnverifiedQ-learning is off-policy learning, which updates Q-values using a greedy policy but can employ different policies such as ε-greedy during exploration.61% similarUnverifiedSPLADE 是一种学习型稀疏检索模型，通过神经网络为文档生成稀疏向量表示，可在倒排索引上高效运行61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/778958API
curl https://kongchang.com/api/v1/knowledge/claims/778958MCP
get_claim(id=778958)