Probabilistic Logic Markov Decision Processes for Modeling Driving Behaviors in Self-driving Cars

Abstract

Rule-based strategies and probability models are among the most successful techniques for selecting driving behaviors of self-driving cars. However, there is still the need to explore the combination of the flexibility and conceptual clarity of deterministic rules with probabilistic models to describe the uncertainty in the spacial relationships among the entities on the road. Therefore, in this paper we propose an action policy obtained from a probabilistic logical description of a Markov decision process (MDP) as a behavior selection scheme for a self-driving car to avoid collisions with other vehicles. We consider three behaviours: keep distance, overtaking, and steady motion. The state variables of the MDP signal the presence or absence of other vehicles in the surroundings of the ego car. Simple probabilistic logic rules characterize the probability distribution of the immediate future state of the autonomous car given the current state and action. The utility model of the MDP rewards the autonomous car when no car is ahead and it penalizes two types of crashes accordingly to their severity. We simulated our proposal in 16 possible scenarios. The results show the appropriateness of both, the overall system and the decision-making strategy to choose actions that prevents potential accidents of the self-driving car.

This work was supported by UNAM-DGAPA under grant TA101222 and Consorcio de IA CIMAT-CONACYT.