Method for controlling behavior diversity of AI intelligent agent in confrontation competitive game

The invention discloses a method for controlling behavior diversity of AI agents in confrontation competitive games, which comprises the following steps of: S1, abstracting a game target diversity behavior mode, and converting the game target diversity behavior mode into a reward function; s2, performing OneHot coding on the diversity behavior mode, and connecting the diversity behavior mode with an existing observed value in series; and S3, constructing a diversity network in parallel with the reinforcement learning backbone network, inputting the observation values after series connection into the diversity network, and summing the output Da of the diversity network and the output Dv of the backbone network as a network code Dt. The method has the advantages that the backbone network and the diversity network are separately arranged, calculation consumption during training of the diversity behavior modes is effectively reduced, and meanwhile, independent training of different diversity behavior modes is avo