Relational reinforcement learning and recurrent neural network with state classification to solve joint attention

摘要

Joint attention is an important non verbal communication learned by humans in a period of childhood. One learning method has been explored to provide this learning ability in robots is known as reinforcement learning. However, the use of this method using a Markov Decision Process model has problems. In this article, we have enhanced our robotic architecture, which is inspired on behavior analysis, to provide to the robot or agent, the capacity of joint attention using combination of relational reinforcement learning and recurrent neural network with state classification. We have incorporated this improvement as learning mechanism in our architecture to simulate joint attention. Then, a set of empirical evaluations has been conducted in the social interactive simulator for performing the task of joint attention. The performance of this algorithm have been compared with the Q-Learning algorithm, contingency learning algorithm and ETG algorithm. The experimental results show that this new method is better than other algorithms evaluated by us for joint attention problem.