Agent training method based on DQN

Abstract

The invention relates to an agent training method based on DQN, which comprises the following steps of: randomly initializing network parameters of DQN networks to obtain a plurality of initial DQN networks; mapping all network parameters of each initial DQN network to obtain an individual, and forming an initial population by all individuals; performing differential evolution operation on the initial population to obtain a new generation of network parameter population, and taking the new generation of network parameter population as a next generation of initial population to repeat the differential evolution operation until a preset evolution condition is reached to obtain a final network parameter population; evaluating each individual in the final network parameter population by using a preset fitness function, and outputting information of an optimal individual; initializing network parameters of the DQN network according to the information of the optimal individual; and training the parameter-initialized DQN network to obtain the intelligent agent. According to the method, in the DQN training process, the execution efficiency is improved, time resources needed by training are reduced, and the instantaneity of intelligent agent training in the game problem is improved.

Description

technical field [0001] The invention belongs to the technical field of artificial intelligence, and in particular relates to a DQN-based intelligent agent training method. Background technique [0002] Deep Reinforcement Learning (DRL), as a hot topic in the field of artificial intelligence at present, integrates the ideas of reinforcement learning and deep learning, and effectively solves the technical problems that cannot be solved by reinforcement learning. For the decision-making problems of continuous action space and continuous state space, deep reinforcement learning uses neural network to efficiently fit the problem of continuous state and action space, and provides a more effective solution for agents to explore complex high-dimensional environments. . [0003] As the pioneering work of deep reinforcement learning, Deep Q Network (DQN) effectively solves the continuous and large model-free problem in the state space. Its main idea is still similar to reinforcement...

AI Technical Summary

Problems solved by technology

In many cases, depending on the specific problem, the agent obtains less effective reward information in the early stage of the algorithm, which cannot effectively accelerate network convergence, and in the case of less effective environmental information in the early stage, blind exploration will lead to Consumes a lot of time resources, thereby reducing the execution efficiency, resulting in the time resources consumed by the training of the agent in the game problem is too long, and the immediacy is poor

Images