Inverse Reinforcement Learning for Generalized Labeled Multi-Bernoulli Multi-Target Tracking

The work presented here is a preliminary study on the feasibility for using the output of a Generalized Labeled Multi-Bernoulli filter as inputs to online imitation learning via Deep Inverse Reinforcement learning with the end goal of predicting the next states of each trajectory output from the filter. This work samples the labeled state trajectories from the filter and discretizes them to create episodes for learning. It is assumed the multi-target dynamics are unknown, but they are stochastic and act to maximize some unknown reward function. Because the ultimate goal is predicting the multi-target motion using only observations of the targets, Deep Q-learning is used to learn the dynamics. However, as this algorithm depends on the unknown reward function, Deep Inverse Reinforcement learning is used to learn the rewards. Due to the coupled nature of the learning problems, their solutions are iterated in an alternating fashion and upon reaching convergence, future states can be predicted over a given time horizon. The results are preliminary, and many extensions to this work are outlined.