Online machine learning for hearing-dog robot to search for user with considering user’s life rhythm

Abstract

A hearing dog is a sort of assistance dog for hearing impaired people. Physical touch of the dog can alert the people to important life sounds, such as doorbells, alarm clocks, or fire alarms. Hearing dogs can work to assist the people, but the working number of hearing dogs are insufficient all over the world. Instead of hearing dogs, we have been developing a hearing-dog robot. This robot can estimate a user’s location from his/her life rhythm to notify to him/her of important life sounds. By this method, the robot achieved efficient user searches. This paper aims to realize more efficient user searches. This paper proposes a method that can acquire the life rhythm in an early stage. The life rhythm represented by a prior probability distribution. In this method, the robot observes the user’s location many times in the early stages of online learning to acquire the user’s life rhythm. Also, we employ JS divergence to evaluate the prior probability distribution, and dynamically control a frequency of the observation of the user’s location. The experimental result indicated that the method worked efficiently to search for the user.