利用文意理解與觀察學習遊戲規則

Abstract

將自然語言處理中，將文字敘述轉成操作指令是一個長期的挑戰，以探索此開放問題為目標，本篇論文關注一個含有大量操作性文字描述的電子卡牌遊戲：爐石戰記(Hearthstone)，並試圖理解卡牌上文字指示造成的狀態變化。 此文提出一套框架，利用觀察遊戲狀態(State)變化以及語意理解來建構環境。藉由模擬遊戲過程，產生大量遊戲狀態變化以及遊戲動作(Action)的對應關係，最終對未見過的卡牌上的操作性文字敘述進行行為預測。

In the field of natural language processing, transforming text commands into instructions for machines has always been a challenge. To explore this open problem, we focused on a well-known digital card game, Hearthstone, attempting to understand the card text descriptions and predict the resolved board states. In this work we propose an approach to rebuild the environment model through rule text understanding and game state observing. By simulating the gameplay to produce state transition pairs after each card is played, this algorithm can predict fair amount of game states by reading rules from cards it has never seen before.