Modeling coalition formation for repeated games using learning approaches

摘要

In this paper, we introduce the notion of “weight” to task's capability, and describe the use of case-based learning and reinforcement learning in a coalition formation model when games are repeated. Based on the the notion “weight” we introduce, a weight-based coalition formation algorithm is proposed, but this algorithm can't always generate good coalitions, to supplement this, an randomized weight-based coalition formation algorithm is introduced. However, deciding when to use which algorithm is not such an easy thing, so a notion of “degree of similarity” is defined, through learning, an optimal degree of similarity can be attained to solve the above problem. In a word, we handle the coalition formation problem in a more of machine learning and data driven perspective.