Optimization of an Autonomous Car Controller Using a Self-Adaptive Evolutionary Strategy

Abstract

Autonomous cars control the steering wheel, acceleration and the brake pedal, the gears and the clutch using sensory information from multiple sources. Like a human driver, it understands the current situation on the roads from the live streaming of sensory values. The decision-making module often suffers from the limited range of sensors and complexity due to the large number of sensors and actuators. Because it is tedious and difficult to design the controller manually from trial-and-error, it is desirable to use intelligent optimization algorithms. In this work, we propose optimizing the parameters of an autonomous car controller using self-adaptive evolutionary strategies (SAESs) which co-evolve solutions and mutation steps for each parameter. We also describe how the most generalized parameter set can be retrieved from the process of optimization. Open-source car racing simulation software (TORCS) is used to test the goodness of the proposed methods on 6 different tracks. Experimental results show that the SAES is competitive with the manual design of authors and a simple ES.