Scalable evolvable hardware applied to road image recognition

Evolvable Hardware (EHW) has the potential to become a new target hardware for complex real-world applications. However, there are several problems that would have to be solved to make it widely applicable. This includes the difficulties in evolving large systems and the lack of generalization of gate level EHW. This paper proposes new methods targeting these problems, where a system is evolved by evolving smaller sub-systems. The experiments are based on a simplified image recognition task to be used in a roadway departure prevention system and later in an autonomous driving system. Special concern has been given to improve the generalization of the system. Experiments show that the number of generations required for evolution by the new method can be substantially reduced compared to evolving a system directly. This is with no reduction of the performance in the final system. Improvement in the generalization is shown as well.