Multi-Sensor Cooperative Tracking Using Distributed Nash Q-Learning

Traditional target tracking algorithm has a disadvantage of excessive dependence on the environment model. Thus a multi-sensor cooperative tracking method using distributed Nash Q-learning was proposed. Distributed Nash Q-learning with model-free was firstly described. Then sensor action and reward function were defined, which both are very crucial to the learning. Sensor action was only subjected to angle control, and reward function was given by calculating the trace of one time-step prediction error covariance. Nash tragedy can not be directly calculated, therefore, a probability statistics method using Bayesian inference was used to update the Q function. Simulation of passive tracking merely with angle measurements shows that this algorithm can enhance the adaptation to environment change and the tracking accuracy.