The Dynamic Evaluation Strategy for Evolvable Hardware

Evolvable hardware (EHW) has recently become a highly attractive topic for the Fault-tolerant System design because it offers a way of adapting hardware to different environments. However, it is time-consuming when circuits become complex. According to our research, the most time consuming period in genetic algorithm (GA) is the fitness evaluation. To reduce the time, a new method based on fitness evaluation expansion GA is proposed. The fitness evaluation is divided into two stages by a threshold. When the generation is lower than the threshold, a fitness estimate strategy is introduced to estimate the offspring´s fitness. When really evolving the fitness, a self-adaptive random sampling model is applied to select the output node from the Cartesian Genetic Programming (CGP) array. During the evolution process, the random sampling probability can be adjusted dynamically with the concentration degree of individuals, which can short the evaluation time and accelerate the convergence. Experiments show that this method can obtain about 5 times speedup while getting an ideal circuit.