Dynamic evolution of simulated autonomous cars in the open world through tactics

Abstract

There is an increasing level of interest in self-driving cars. In fact, it is predicted that fully autonomous cars will roam the streets by 2020. For an autonomous car to drive by itself, it needs to learn. A safe and economic way to teach a self-driving car to drive by itself is through simulation. However, current car simulators are based on closed world assumptions, where all possible events are already known as design time. Nevertheless, during the training of a self-driving car, it is impossible to account for all the possible events in the open world, where several unknown events may arise (i.e., events that were not considered at design time). Instead of carrying out particular adaptations for known context events in the closed world, the system architecture should evolve to safely reach a new state in the open world. In this research work, our contribution is to extend a car simulator trained by means of machine learning to evolve at runtime with tactics when the simulation faces unknown context events.