TRAINING AND SIMULATION OF NEURAL NETWORKSFOR WEB PROXY CACHE REPLACEMENT

摘要

Web proxy caches are widely used to reducernuser-perceived latency and link congestion causedrnby the extremely high volume of web traffic. Inrnthis research, neural networks are trained to makernproxy cache replacement decisions. The neuralrnnetworks are trained to classify cacheable objectsrnfrom real world data sets using information knownrnto be important in web proxy caching, such asrnfrequency, recency and size. The networks arernable to obtain correct classification ratios ofrnbetween .85 and .88 both for data used for trainingrnand data not used for training. In simulation, thernfinal neural networks achieve hit rates that arern86.60% of the optimal in the worst case and 100%rnof the optimal in the best case. Byte-hit rates arern93.36% of the optimal in the worst case andrn99.92% of the optimal in the best case.