《1995 High-Performance Job-Shop Scheduling With A Time-Delay TD(lambda) Network》.pdf

HighPerformance JobShop Scheduling With A TimeDelay TD  Network Wei Zhang and Thomas G Dietterich Department of Computer Science Oregon State University Corvallis Oregon fzhangw tgd gresearchcs orstedu Abstract Jobshop scheduling is an imp ortant task for manufacturing indus tries We are interested in the particular task of scheduling payload pro cessing for NASAs space shuttle program This pap er summa rizes our previous work on formulating this task for solution by the reinforcement learning algorithm T D A shortcoming of this previous work was its reliance on handengineered input features This pap er shows how to extend the timedelay neural network TDNN architecture to apply it to irregularlength schedules Ex p erimental tests show that this TDNNT D network can match the p erformance of our previous handengineered system The tests also show that b oth neural network approaches signicantly out p erform the b est previous nonlearning solution to this problem in terms of the quality of the resulting schedules and the numb er of search steps required to construct them Category Control Navigation and Planning Reinforcement Learning Presentation Preference Poster Intro duction In Tesauros landmark work on TDgammon he showed that the temp oral dierence algorithm T D Sutton can learn an excellent evaluation function for the game of backgammon This is probably the most successful application of reinforcement learning to date The goal of our research is to determine whether this success can b e duplicated in