원문정보
초록
영어
Prisoners’ dilemma is a typical game theory issue. In this study, it was treated as an incomplete information game to establish a related machine learning model using a naive Bayesian classification method. The model established was referred to as the Bayes model. Using this model, the incomplete information game was soluble with the assistance of statistical machine learning. This study proceeded as follows: firstly, four typical models were run against the Bayes model some 10,000 times. The total incomes of the models recorded suggested that Bayes model was more advantageous than other models. Even in a multi-player prisoners’ game, Bayes model also presented the desired level of performance and accrued a higher income than other models. Further statistical analysis implied that the Bayes model and the widely accepted optimum strategy tit-for-tat (TFT) model showed a tendency to be prone to defection. Secondly, according to the games run on the natural Bayes model, as well as the natural TFT model, it was found that the Bayes model accrued more benefits than the TFT model on average. Finally, comparison of the Bayes model with the TFT model revealed that the Bayes model was better. This demonstrated the efficacy of the Bayes model constructed in this study and moreover, provided a novel idea for solving the problem of an incomplete information game.
목차
1. Introduction
2. Construction of the Model
2.1. Prisoners’ Dilemma Model
2.2. Typical Strategy Models
2.3. The Bayes Model
2.4. The Multi-player Prisoners’ Dilemma Model
2.5. Strategy in the Multi-player Prisoners’ Dilemma based on Statistical Machine Learning
2.6. Evaluation of the Strategy Model
3. Experimental Results and Analysis
3.1. The Performance of the Double-player Strategy Model: Naive Bayesian Classification
3.2. The Performance of the Multi-player Bayes Model
3.3. Analysis of the Performance of the Bayes Model versus Common Models
3.4. The Performance of the Bayes Model when Run over Fewer Games
3.5. The Game Results from a PTFT Model Compared with the Other Models
4. Conclusions
Acknowledgements
References