A Probabilistic Decision Engine for Navigation of Autonomous Vehicles under Uncertainty

Navigation or route planning under uncertainty is a very challenging but important task for autonomous vehicles such as self-driving cars, drones, unmanned aerial systems (UAS), etc. Probabilistic methods to the modeling and optimization of these systems are attractive in quantitatively capturing the uncertainty present in their dynamic environment. This paper will present a novel probabilistic decision engine that can serve as the core for the navigation of autonomous vehicles under uncertain conditions. This probabilistic decision engine takes in a network connection matrix (based on maps and graph theory) and a cost matrix (with entries of the cost’s mean values and probability distributions) as its input and generates the probability distributions of the optimal routes as its output. The proposed probabilistic decision engine consists primarily of a stochastic network standardization module, a stochastic network decomposition module and a probabilistic solver (i.e., decision kernel). A deterministic network reduction method based upon Dijkstra’s algorithm is first used to derive a standard, reduced network, augmented by a stochastic network reduction process. The standard network is then decomposed into a series of stochastic subnetworks by using sequential convolution, PDF (probability distribution function) shifting and reshaping techniques. A purely-analytical probabilistic solver is finally used to solve the stochastic decision-making problem. In this paper, the principle of operation and implementation methods of the entire probabilistic decision engine will be discussed in detail. Some representative simulation results will be provided to demonstrate the effectiveness of the proposed computational methodology and compared with the traditional Monte-Carlo simulation method to validate the analytical results. This study suggests that the time needed to find the solution using the proposed decision engine can be reduced by three to four orders of magnitude, compared with the Monte-Carlo simulation method.