Unverified50% confidenceFactExact time
DQN的经验回放缓冲区通常存储100万条(s,a,r,s')元组，目标网络每隔固定步数（如10000步）同步参数
1
Sources
50%
Confidence
Long-term
Relevance
8/21/2026
First Seen
Sources
Related Claims
Unverified目标网络维护参数更新滞后的网络副本来计算目标值，每隔固定步数（如1000步）将主网络参数同步过去，解决追逐移动目标问题78% similarUnverifiedReAct's reasoning chain continuously consumes context tokens; for an agent executing 20 steps, the accumulated reasoning history may consume tens of thousands of tokens.62% similarUnverifiedAnthropic provides a prompt caching mechanism whereby, when the system prompt remains unchanged across consecutive calls, the tokens in the repeated portion are billed at only 10% of the first-call cost.62% similarUnverified多智能体系统的Token优化策略包括模型路由、压缩智能体间传递的上下文、缓存中间结果62% similarUnverifiedAnthropic提供的提示词缓存机制允许将重复出现的长系统提示或技能定义缓存于服务端，后续调用仅需传输差异部分61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/782691API
curl https://kongchang.com/api/v1/knowledge/claims/782691MCP
get_claim(id=782691)