Improving the learning speed in reinforcement learning issues based on the transfer learning of neuro-fuzzy knowledge

This paper to the topic of transfer learning in environments that share some of its features. The main challenge in this topic is how to transfer knowledge from the source environment to the target environment. In the presented idea, taking into account the common features in the operating space between the two environments, the value of the operation in the source environment first is obtained and then it uses a neuro -fuzzy network to approximate the value of the value function of the operation. In the target environment, the value of the mode of operation is used to combine the predictive value of the neuro - fuzzy network and the amount received in the environment itself. In other words, according to the training carried out in the source environment, value-action values in the target environment are derived from the combination of value-action values approximated by the neuro - fuzzy network and the amount obtained from the learning algorithm in that environment. It is worth noting that the learning algorithm Q is used in the environment. The results of the proposed idea indicate a significant increase in learning speed.