干旱区生态环境知识资源中心

This paper uses GAssist system to get symbolic rules and proposes four techniques to improve it. A new population initialization method is applied and fitness scaling is used to promote the population's convergence. It also improves on the deserted hierarchical selection operator and combines it with the MDL-based fitness function to control bloat effect. Finally, a new stop criterion to GA is studied. The experimental results show that the system is further improved. Comparing with other systems GA system is roughly comparable in generalization capacity but the efficiency needs improve.