運用深度學習預測中華職棒比賽勝負:以中信兄弟象隊為例

ASIA unversity > 資訊學院 > 資訊傳播學系 > 博碩士論文 > Item 310904400/112477

Please use this identifier to cite or link to this item: http://asiair.asia.edu.tw/ir/handle/310904400/112477

Title:	運用深度學習預測中華職棒比賽勝負:以中信兄弟象隊為例 Prediction of Match Results for Chinese Professional Baseball League Using Deep-Learning Approach: A Case Study of ChinaTrust Brothers Elephant Team
Authors:	蔡育楷 TSAI, YU-KAI
Contributors:	資訊傳播學系
Keywords:	深度學習;類神經網路;職棒;棒球;勝負預測 deep-learning;neural networks;major baseball;baseball;victory and defeat prediction
Date:	2019
Issue Date:	2019-11-11 03:48:18 (UTC+0)
Publisher:	亞洲大學
Abstract:	深度學習神經網路在棒球上的運用，目前相關研究所佔比例極少，而棒球比賽最受關心的部份是最後的勝敗結果，也是球迷最關心的事情之一。我們既有的觀念中，類神經網路訓練需仰賴大量的數據及訓練資料才能提高準確度，但教練的指揮調度及球員的狀態起伏，常隨著球賽累積經常調整而有所調整，所以面對賽季長且場次多的職業棒球來說，以大數據訓練神經網路的方式未必合適。本研究利用深度學習方式創建類神經網路模型，預測中華職棒大聯盟比賽勝負，並且以中信兄弟象隊例行賽為例，我們使用的特徵參數包括：對戰兩隊的團隊勝敗戰績、打擊率、自責分率及先發投手的出賽數、勝場數、防禦率、被打擊率、每局被上壘率…等共165項，將這些特徵參數饋入深度學習神經網路之後，判斷最後輸出結果，輸出結果有3類，分別判定為勝、敗、和。經由實驗結果證明，使用深度學習神經網路確實可以提供預測比賽勝負的參考依據，其中以比賽前10場比賽作為訓練資料預測準確度最高，預測正確率達到60%，優於各預測模型，訓練數據過多過少皆會影響預測模型之準確率。 Deep learning neural networks are widely developed and can be applied in the analysis of a baseball game, where the match result of the baseball game is interested to the fans. Neural network training relies on a large amount of data to improve accuracy, but the match result also depends on coach's command and players’ statuses. These factors are complex and time variated. Accordingly, training neural networks with big data may not be appropriate for a long competition season. This study attempts at utilizing deep-learning neural networks to predicting the outcome of the China Major League Baseball game, and taking the ChinaTrust Brothers Elephant Team as an example. The features we used include: the match result records, strike rate, self-blame rate, the number of wins, defensive rate, the rate of attack, the rate of hits, etc., a total of 165 items. By feeding these features into the deep-learning neural networks, a predicted match result can be obtained. The match results are classified into three categories, including victory, defeat, and tie. The experimental results show that the classified result of deep-learning neural networks can provide a valuable reference for predicting the outcome of the game. The training features obtained by the nearest ten games provide the highest prediction accuracy, which can reach more than 60%.
Appears in Collections:	[資訊傳播學系] 博碩士論文

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	322	View/Open

View Licence

Loading...