Behavior evolution-inspired approach to walking gait reinforcement training for quadruped robots