Travel Time-Dependent Maximum Entropy Inverse Reinforcement Learning for Seabird Trajectory Prediction

摘要

Trajectory prediction is a challenging problem in the fields of computer vision, robotics, and machine learning, and a number of methods for trajectory prediction have been proposed. Most methods generate trajectories that move toward a goal in a straight line (goal-directed) while avoiding obstacles. However, there are not only such goal-directed trajectories but also trajectories that taking detours to reach the goal (non-goal-directed). In this paper, we propose a method of predicting such non-goal-directed trajectories based on the maximum entropy inverse reinforcement learning framework. Our method introduces travel time as a state of the Markov decision process. As a practical example, we apply the proposed method to seabird trajectories measured using global positioning system loggers. Experimental results show that the proposed method can effectively predict non-goal-directed trajectories.