Decision making rule extraction from action logs

Abstract

To analyze agents' decision-making process or rules about their actions, a long period of time is required for researchers to observe the target agents' actions. For example, in order to use such agents' actions as training data for artificial agents, numerical data and its linguistic expressions are needed. In this research, first, neural network agents are trained using numerical data from the log of RoboCup. Next, decision-making rules are extracted from the input-output relations of the trained neural network agents. Experimental results show the effectiveness of the proposed technique.