Title: 非監督式模糊神經網路推理模式於個人電腦叢集之平行計算
Parallel Computing of Unsupervised Fuzzy Neural Network Reasoning Model on PC Cluster System
Authors: 謝景惠
Ching-Hui Hsieh
Shih-Lin Hung
Keywords: 平行計算;個人電腦叢集;Parallel Computing;PC Cluster;MPI
Issue Date: 2001
Abstract: UFN推理模式(Unsupervised Fuzzy Neural Network Reasoning Model)已被證實是一套在解決工程問題上相當有效率的人工神經網路(ANN)模式。案例庫案例數目的增加可增進UFN推理模式的預測精確度,但卻會因此而增加計算時間以達到要求的精度。而近年來,已證實在透過訊息傳遞介面(Message-Passing Interface, MPI)建立平行環境的個人電腦叢集(PC Cluster)上進行平行計算(Parallel Computing)對於解決需要大量計算時間的問題也是一個相當合適的選擇。因此在本研究中,提出平行化的UFN推理模式並將之透過C語言加上MPI的函式庫在個人電腦叢集上進行平行計算,並利用美國鋼筋混體土規範ACI 318-95中對樑的抗剪力公式產生數據,評估此一平行UFN推理模式的計算效能。在一套由兩台雙CPU的電腦共四顆Intel PⅢ-500所組成的的個人電腦叢集上進行此一平行化UFN推理模式可達到平均約3.7的速度提升值,此外在15000組訓練案例下,5000組驗證案例的預測結果平均相對誤差可收斂至2.63%。
Unsupervised Fuzzy Neural Network (UFN) reasoning model has been proved as an effective ANN model for solving engineering problems. The performance of the model may enhance as the number of training instance increases; however, it needs more computation time to reach desired accuracy. Recently, it is also proved that parallel computing via PC cluster with MPI parallel environment is a feasible solution to solve problems with much computation time. In this research, a novel parallel UFN model is developed and implemented in C language with an MPI library on PC cluster. The computational performance of the novel model is also investigated via a beam design problem with specifications of ACI 318-95. Average 3.7 speedup is achieved on the PC cluster with two nodes of four Intel PIII-500 Mhz CPUs. Moreover, the average relative error of 15000 training instances is converged to 2.63 percent.
Appears in Collections:Thesis