Dynamic Game Theoretic Electric Vehicle Decision Making

Abstract

<div>Real-world driving in diverse traffic must cope with dynamic environments including a multitude of agents with uncertain behaviors. This poses a challenging motion planning and decision-making problem, as suitable algorithms should manage to obtain optimal solutions considering nearby vehicles. The state-of-the-art way of environment and action generalization is built on mathematical modeling and probabilistic programming of idealistic incidents. In this article we present dynamic anytime decision making, a decision-making algorithm that takes advantage of natural evolutionary and developmental processes to make decisions for an autonomous vehicle navigating in traffic. The methodology to achieve multidimensional judgment under unforeseen circumstances is to enable stochastic Bayesian game theory when modeling interactive properties and scenario estimation. In particular, we employ a dual-layer optimization strategy that substantially enhances the capability to extract mutual influence into decision matrix that organically synthesize decision rules according to encountered situations. Furthermore, our decision-making algorithm incorporates incremental capability that computes modified control variable due to unexpected vehicle behavior or intention. This results in improved performance gain where dynamic conditions receive reasoning with more effectiveness. We produce swift policy correction by leveraging previous game mechanism, where cooperative characteristics of both the controlled vehicle and other agents are integrated without computation overload. Aiming to reach utmost accurate decision, our reward function is reinforced with weight-tuning function to deduce optimal hyper parameters in a wider spectrum of vehicle scenarios. Through coupled decision tree optimization and incremental trajectory control, economical commands are computed with enhanced precision under complex traffic pattern. To this end, our system provides a robust framework in terms of handling dynamic vehicle interactive control, as decision is computed by synergy between progressive interface and external rationalization. We demonstrate the algorithm performing preeminent passing maneuvers and merging in a number of different simulated scenarios.</div>