標題: 應用類神經網路探討火災保險風險評估之研究--以T公司工商火災保險為例
Study of the Application of Artificial Neural Network Research on the Risk Assessment on the Fire Insurance:Case Study
作者: 孫其怡
Chi Yi Sun
陳春盛
Chun Sung Chen
工學院產業安全與防災學程
關鍵字: 火災保險;火災風險評估;類神經網路;fire insurance;fire risk assessment;artificial neural network
公開日期: 2007
摘要: 摘 要 全世界各國產險市場發展的趨勢已逐漸走向費率自由化,在產物保險發展的現況中,費率自由化便成為保險監理的既定政策,未來火災保險因費率自由化,各公司將會面臨更多的挑戰,尤其是對風險的評估能力更是重要,是否有一套完善的風險評估標準,來決定是否承保或者是以什麼條件來承保,實在是各產險公司火險核保部門或是風險評估部門的當務之急。 本研究藉由人工智慧中的類神經網路(Artificial Neural Network, 簡稱ANN)建立火災保險風險評估模型,來輔助產險核保風險評估決策。本研究主要彙集前人的研究與相關評價模式,再利用保險公司承保資料與以往客戶出險的紀錄,進而建立火災保險風險評估的項目,以期能建立一套適用於本土化的風險評估原則。 研究的實驗結果摘要如下: 1.最佳化參數選擇:透過多次的參數敏感度實驗,以神經元個數 = 10、學習率 = 0.6、慣性因子 = 0.6 所建構倒傳遞類神經網路,可訓練出整體正確率最高且失誤偵測個數相對較低的權重值,以此權重與參數值來建構倒傳遞類神經網路,作為未來測試樣本之模組。 2.正確率:本研究之訓練樣整體正確率可達到 85.50%;測試樣本的整體正確率可達到 82.42 %,可以推斷該類神經網路結果應具有一般性與可信度。
Abstract The current trend of products insurance market is towards open competition instead of tariff rate in the world. In domestic products insurance market moves towards competition market gradually. In the future, all insurance companies will face more challenges in fire risk assessment especially. All insurance companies should have a set of perfect risk assessment or standard whether or not. The paper tries to establish the fire insurance risk assessment model by the artificial neural network (ANN). The purpose of the research is to assist us in fire risk assessment decision-making. This paper mainly collects predecessor's research and the items of the correlation appraisal models. Furthermore, using customers’data of some insurance company to establish fire insurance risk assessment model. We want to establish a model to suit for the localization risk assessment principle. The research experimental result is as follows: 1.Optimization parameters choicing: While we experiment with parameters using try and error method to try finding optimization parameters. We get optimization parameters to establish a model by the neuron numbers of the hidden layer = 10, the learning rate = 0.6, the momentum factor = 0.6.The model may prepare for testing samples in the future. 2.Accuracy rate: In the model, we use the training samples to achieve 85.50% accuracy rate. Using the testing sample overall accuracy rate may achieve 82.42%. We may infer the artificial neural network to establish the model supposed to have its generality and confidence level.
URI: http://140.113.39.130/cdrfb3/record/nctu/#GT009366523
http://hdl.handle.net/11536/80056
Appears in Collections:Thesis


Files in This Item:

  1. 652301.pdf
  2. 652302.pdf