Behavior Decision Of Game Agent Based On Artificial Intelligence

In the aspect of game agent behavior control,there are many artificial intelligence as control methods in the world.Open AI research results show that the specific evolutionary strategy is better than the original training effect based on gradient descent method when training large neural networks,and correspondingly requires more computational power.On the application level,this paper attempts to propose a logic behavior control method for game agents with low computational power,i.e.an algorithm that uses reinforcement learning to adjust the structure of Shenjingluo on the basis of neural network training based on genetic algorithm.This paper constructs a multi-agent environment in Unity3 D operating environment,in which training tasks are divided into two phases.Firstly,in the algorithm of training neural network based on multi-strategy genetic algorithm.Using genetic algorithm as search method has good parallelism and strong global search ability.However,the disadvantage is that when the population is not large enough,it is easy to reduce the overall gene abundance of the population due to too strong individual gene competitiveness in the early training period,and eventually fall into local optimization.The use of multi-strategy genetic scheme can effectively reduce this shortcoming.Secondly,in the neural network algorithm based on reinforcement learning.The structure of the neural network is marked as a state in reinforcement learning,and the network structure is adjusted once every training period.Since the network information will not be lost in the process of changing the network structure,thus interrupting the training process,the period of adjusting the network structure can be very short,thus improving the training efficiency of the network,and at the same time further reducing the possibility of premature entry into local optimization caused by genetic algorithm.The purpose of this paper is that under the condition of low computational power,the neural network using reinforcement learning and genetic algorithm dynamic adjustment can still meet the requirements of game agent behavior logic control.The experimental results show that the game agent trained by the multi-strategy genetic algorithm can complete the basic behavior logic,and after adding the reinforcement learning adjustment neural network module,the fitness of the game agent is obviously improved,and more complex behavior logic can be completed at the same time.