Autonomous Vehicles Through the Prism of the Turing Test

Abstract

In recent years, significant progress has been made in research on various problems of autonomous transport. According to most forecasts, autonomous vehicles will appear on the roads in the coming years. However, strict requirements for the artificial intelligence of autonomous vehicles have not yet been formulated. We do not yet have a clear understanding of the intelligence for autonomous vehicles. Nevertheless, the problem of developing an analogue of the Turing test for autonomous vehicles has attracted increasing attention in recent years. There are a number of different points of view on the analogue of the Turing test for autonomous vehicles. We show that passing the Turing test must be performed under conditions that are significantly different from those commonly used. We argue that passing such a test can be presumably much harder than the original. We consider a number of additional tests that can be used as some parts of the Turing test. In particular, we can mention such tests as practical test of driving skills, health test, test of prediction skills, security test, body test, coexistence test, aging test, trust test, overall performance test, no-harm test. In this paper, we pay special attention to the no-harm test. We consider an approach that is based on evolutionary machine learning. For the first study of the no-harm test, we have considered a relatively simple model of the natural environment and have proposed an algorithm for artificial evolution for the environment. The results of our experimental studies show that insufficiently justified implementation of autonomous vehicles can lead to unpredictable consequences.