基于增强学习的半导体测试调度研究  

Scheduling Study for Semiconductor Final Test Based on Reinforcement Learning

作　　者：张智聪[1] 郑力[2] 翁小华 

机构地区：[1]广东东莞理工学院工业工程系,东莞523808 [2]清华大学工业工程系,北京100084 [3]南佛罗里达大学工业与管理系统工程系,佛罗里达坦帕33620

出　　处：《工业工程与管理》2009年第4期38-44,59,共8页Industrial Engineering and Management

基　　金：国家自然科学基金(70771058);国家自然科学基金(50375082);国家863计划资助项目(2008AA04Z102)

摘　　要：采用Sarsa(λ,k)学习算法求解、产品、测试机、测试工具包、使能器部件对应关系非常复杂的半导体测试调度问题。针对测试调度,通过定义系统状态的表示方式、构造行为和报酬函数把调度问题转化为增强学习问题,并把Sarsa(λ,k)算法和梯度下降径向基神经网络函数泛化器结合使用。实验验证了Sarsa(λ,k)算法解决半导体测试调度问题的有效性。Sarsa(λ,k)算法通过反复解决调度问题来调整调度策略,能克服单个行为策略短视的缺点,综合利用各个行为策略的优点,从而找到较优的调度方案。Semiconductor test scheduling problem is a variation of reentrant unrelated parallel machine problem considering intricate multiple resources constraints and sequence-dependant setup times, etc. A multi-step reinforcement learning（RL）algorithm called Sarsa（λ, k）was applied to deal with the semiconductor final test scheduling problem. Allowing enabler reconfiguration,the production capacity of the test facility was expanded and scheduling optimization was performed at the component level. In order to apply Sarsa（λ, k）, the scheduling problem was transformed into an RL problem by defining state representation, constructing actions and the reward function, and combining the algorithm with the gradient descend radial basis neural networks function approximation. Experiments show that Sarsa（λ,k） outperforms the scheduling method in industry and validate its effectiveness to solve the semiconductor test scheduling problem.

关 键 词：调度 半导体测试 增强学习 多资源约束 

分 类 号：TP311.52[自动化与计算机技术—计算机软件与理论]