本研究採用「強化式」機器學習策略，跳過建構股價漲跌預測系統，而直接建構遺傳神經網路 (Genetic Neural Networks, GNN) 買賣決策系統，並針對台灣股市實證幾個重要的課題。研究結果顯示：(1) 在大盤指數方面，在GNN的演化過程中，確實可以觀察到「訓練期間績效與測試期間績效相關」與「隨著演化世代增加績效增加」現象，顯示GNN確實學習到具普遍化獲利能力的大盤交易策略。(2) 使用包含成交量的資訊產生的系統如果能避免過度學習，可以提高投資績效。(3) 使用短訓練期間 (4.5年) 的系統的獲利明顯小於使用長訓練期間 (12年) 者，顯示4.5年的訓練期間太短，不足以學習到具普遍化獲利能力的交易策略。(4) 使用「多遺傳神經網路多數決策略」顯示採用多數決策略無助於提高對大盤的投資績效，但可使其更穩定。(5) 在類股指數方面，其獲利能力等同買入持有策略，顯示GNN決策系統無法提高類股投資績效。

This study employed "Reinforced Learning" strategy to bypass stock price f1uctuationprediction stage and construct stock trading decision system using Genetic Neural Networks (GNN) directly. The system was validated by several important topics aiming at Taiwan stock market. The results showed the following conclusions. (1) In the evolution process of GNN with regard to stock index of Taiwan, two phenomena can be observed. First, the training period performance is correlated with test period performance. Second, the performance increases with each evolution generation of GNN. These two phenomena demonstrated that GNN can learn the general profitable trading strategy on stock index of Taiwan. (2) The trading system using price as well as volume information could increase investment perform­ance if it can avoid over-learning. (3) The profit of the trading system using short period in­formation (4.5 years) is obviously smaller than that using long period information (12years). It demonstrated that 4.5 years is t oo short to learn the general profitable trading strategy. (4) Using "majority decision strategy based on multi-GNNs" can not increase the mean but can reduce the standard deviation of profit. It demonstrated that this strategy is useful to improve the stability of investment performance on Taiwan stock market. (5) With regard to sector index, the profit of the trading system is about the same as the buy-and-hold strategy. It demonstrated that the system can not increase the investment performance on the sector index.