“Credit assignment”指在一个结果(成功或失败)出现后,确定哪些行动、决策、步骤或因素应当为该结果“记功”或“担责”的过程。该术语常见于机器学习与强化学习中,尤其指如何把奖励/误差合理分配到先前的行为或网络参数上。(也可泛指组织管理中对贡献的归因与分配。)
/ˈkrɛdɪt əˈsaɪnmənt/
In reinforcement learning, credit assignment helps the agent learn which actions led to a reward.
在强化学习中,信用分配帮助智能体学会哪些行为带来了奖励。
Because the effects were delayed and indirect, the credit assignment problem became difficult to solve.
由于影响是延迟且间接的,信用分配问题变得很难解决。
“Credit”原意与“相信、信任”相关,后来引申为“功劳、信誉、学分”等;“assignment”来自“assign(指派、分配)”。合起来字面意思是“把功劳分配出去”,在科学语境中进一步抽象为“把结果的因果贡献归因到先前步骤/变量上”的问题。