A taxonomy for similarity metrics between Markov decision processes

Abstract Although the notion of task similarity is potentially interesting in a wide range areas such as curriculum learning or automated planning, it has mostly been tied to transfer learning. Transfer based on idea reusing knowledge acquired set source tasks new process target task, assuming that and are close enough . In recent years, succeeded making reinforcement (RL) algorithms more efficient (e.g., by reducing number samples needed achieve (near-)optimal performance). RL core concept : whenever similar , transferred can be reused solve significantly improve performance. Therefore, selection good metrics measure these similarities critical aspect when building algorithms, especially this from simulation real world. literature, there many between MDPs, hence, definitions its complement distance have considered. paper, we propose categorization analyze proposed so far, taking into account categorization. We also follow taxonomy survey existing well suggesting future directions for construction metrics.