標題: 在即時戰略遊戲中探勘與預測玩家的策略風格
Mining and Predicting the Strategy Style of Player in Real-Time Strategy Game
作者: 王正宏
Zheng-Hong Wang
孫春在
Chuen-Tsai Sun
資訊科學與工程研究所
關鍵字: 即時戰略遊戲;策略風格;決策探勘;代理人自動決策系統;Real-Time;Strategy;RTS;case-based reasoning;decision-making system
公開日期: 2006
摘要: 如何設計代理人的決策系統,並使其能夠自動學習或發展新的策略一直是人工智慧領域尚待解決的問題。本論文在即時戰略遊戲(Real-Time Strategy Game)的環境下,藉由分析玩家的策略風格以建立具學習能力的決策系統。除了建立具有特殊玩家風格的自動決策系統外,尚能做為假想敵以供對戰演練。即時戰略遊戲的策略考量包括了資源分配、資訊隱藏、善用環境以及策略佈局,這些在遊戲中的考量權衡與真實世界中的戰場、商場、以及球場上所考慮的主要因素是一致的。在即時戰略上設計有計畫與學習能力的決策系統常會遭遇到複雜性的問題,例如決策空間太過龐大,系統不太可能即時下定決策,部分研究者嘗試以人工設計幾種固定策略,再從中採用動態挑選、學習的機制,這種作法雖然成功的降低了策略空間的複雜度,卻又回到了最原始的問題-「電腦的策略(打法)在一定的程度上被預先限制住了」。 人類之所以能夠進步快速的原因就在於懂得善用前人的經驗,生病的時候我們知道要吃什麼藥,而不需要再嘗百草。 透過過去發生的記錄,人類可以預知什麼行為可能會帶來什麼結果,我們學會如何去擇優避劣而不需要親自體驗。 即時戰略所儲存的記錄檔是許多玩家競賽後留下的寶貴經驗,然而目前的即時戰略人工智慧卻缺乏吸收這些經驗的設計。 在這次研究當中我們會收集大量的記錄檔,然後根據我們想要了解的玩家作篩選,篩選的方式是透過玩家的個人習慣,並將篩選出來的資料作為輸入,探勘出玩家在面臨各種狀況下的決策,以瞭解並形成玩家的個人風格。 我們所採用的核心方法是案例推論(Case Based Reasoning),已經有許多研究者很成功的將這套方法運用在各種領域,像是醫學、遊戲、以及機器人的移動上。 實驗的最後我們比較決策系統所做的決策,以及記錄檔中玩家的實際決策,來驗證我們的模型。 在驗證的過程中我們發現系統很容易達到五成以上的預測準度,從一場場比賽的預測準度變化來看,不同的玩家在遊戲中的策略方針也的確有各自的特色。
The author built a decision-making system with learning ability in real-time strategy(RTS) games. The focus falls on the strategy style of player which is used to build an intelligent opponent agent for practicing. The factors of strategy considered in RTS games include resource management, information hiding, environment, and strategy priority-arrangements. Similar practical experience and application can be found in battlefield, business or ball game. Most of decision system in real-time strategy games with planning and learning ability cannot avoid uncertain and unpredictable complexity problem— “It cost too much time to make a decision in a complex, real-time environment." Some researchers try to dynamically choose and evolve sub-plans between fixed handmade plans. Although this kind of design was successful in reducing time complexity, it was limited in a sense. In this thesis, case-based reasoning approach is applied to analyzing gaming experience from player’s playing log for building a decision-making system. The end of this thesis verifies the proposed model by comparing the strategy made by our system and real decision made by human players. We found that each player has formed significant personal strategy style which can be found in different game plays.
URI: http://140.113.39.130/cdrfb3/record/nctu/#GT009455543
http://hdl.handle.net/11536/82066
Appears in Collections:Thesis


Files in This Item:

  1. 554301.pdf

If it is a zip file, please download the file and unzip it, then open index.html in a browser to view the full text content.