Interactive Q-learning on heterogeneous agents system for autonomous adaptive interface

Purpose of this system is to adapt the bedridden people who cannot move their body easily, so the simple reinforcement signals are applied. The application is to control the behaviors of Khepera robot, which is a small mobile robot. For the simple reinforcement signals the on-off signals are employed when the operators as the training agent feels discomfort for the behaviors of the learning agent Khepera robot. We proposed the new reinforcement learning method called Interactive Q-learning and the heterogeneous multi agent system. Our multi agent system has three kinds of heterogeneous single agent: Learning agent, Training agent and Interface Agent. The system is hierarchic. There are also three hierarchies. It is impossible to iterate the many episodes and steps to converge the learning which is adopted in general reinforcement learning in simulation world. We show the results of experiments using the Khepera robot for 3 examinees, and discuss how to give the rewards according to each operator and the significance of heterogeneous multi agent system. We confirmed the effectiveness through the some experiments which are to control the behavior of Khepera robot in real world. The convergences of our teaming system are quite quick. Furthermore the importance of the interface agent is indicated. The individual differences for the timing to give the penalties are happened even though all operators are young.