Research And Application Of Transfer Model Based On Relational Mapping

With the development of machine learning,transfer learning has attracted the attention of more and more researchers.The reason that transfer learning is valued is that it can play an important role in improving the adaptive ability of the algorithm,solving the problem of insufficient data in the field and reducing the training time of the algorithm.Transfer learning introduce source domain to solving the original problem which is called target domain.If the data distribution,feature space and output space in source and target domain are all the same,transfer is homogeneous.Conversely,if they are not all the same,transfer is heterogeneous.This means that transfer learning breaks the i.i.d.assumptions of traditional machine learning,so learning will be more difficult.This paper studies heterogeneous transfer learning.We transfer from source domain to target domain by using relational mapping.Its essence is to construct the relationship between source domain and target domain and then use this relationship to transfer.In the area of transfer learning,this is a relational-based transfer learning.We combined the transfer method based on relation mapping with two different machine learning methods and obtained the transfer Markov logic networks and transfer reinforcement learning.Markov logic network(MLN)is a machine learning model that use logical formulas to infer.We use predicate mapping to construct formulas of target domain from source domain.We select mapping formulas by rule filter and predicate covering.We tested our transfer learning algorithm on three public data sets.The results show that our algorithm is better than the existing two transfer MLN algorithms.In addition,we applied the transfer algorithm on OSN data sets,successfully established a MLN with a small amount of data in target domain and had a good inference effect.Reinforcement learning widely use in the robot field and it usually takes a long training time.We propose a framework of transfer reinforcement learning based on bidirectional mapping.Through two relational mappings,the knowledge in source domain can be used to guide the target domain to shortening the training time in target domain.We apply the framework to Keepaway and propose two transfer learning methods based on the framework.We compare two transfer methods with the best reinforcement learning methods we know in the Keepaway field.The results show that one of our transfer learning methods is better than that transfer learning method.