Considering a method for generating human mobility model by reinforcement learning

Abstract

Reinforcement learning is a machine learning framework that an agent repeatedly observes its reward from environment as a result of action to choose with high reward expectations. On the other hand, mobility models that satisfy statistical properties of human mobility patterns have been proposed. However, there is little study to consider a reasonable method of generating a mobility model based on reinforcement learning. In this paper, we proposed a method for generating a mobility model using reinforcement learning to earn rewards most efficiently. A dataset containing mobility traces of taxi cabs in San Francisco was used to propose the method where each taxi agent learns its actions selected by giving rewards, such as passenger fare, gasoline cost, and the distribution of the positions of past passengers. The (Formula Presented)-greedy method was used to consider the method to perform reinforcement learning. The degree of learning was compared by changing the exploration parameter (Formula Presented). As a result, it was found that the cumulative reward was the highest when (Formula Presented), which is different from usual results in (Formula Presented)-greedy method. This result come from an implicit exploration of taxi’s mobility patterns where taxi agents explored the places where many passengers can be found during the transfer of passengers.