Safe deep reinforcement learning method based on abstract training and verification

The invention discloses a safe deep reinforcement learning method based on abstract training and verification, and the method comprises the steps: state abstraction: abstracting an infinite continuous state space of a reinforcement learning environment into a finite discrete state space according to a preset abstract granularity; training in the abstract state, adjusting a neural network and a loss function in the deep reinforcement learning system to realize training in the abstract state, and acting an output action of the neural network on the environment to obtain a subsequent abstract state; performing formalized security property verification, and checking whether the deep reinforcement learning system meets the security property or not by using action-based calculation tree logic; performing counter-example refinement, further performing subdivision refinement on an abstract state to which counter-examples generated in the formalized security property verification process belong, and performing a train