Reinforcement learning model optimization method and device based on parameter noise

Abstract

The invention provides a reinforcement learning model optimization method based on parameter noise. The parameter noise of periodic Gaussian distribution is added to the parameters of the model in a heuristic mode, the exploration range of unknown decisions in the model training process is enlarged, more excellent decisions can be found more quickly in the model training process, the convergence speed is increased accordingly, and an optimal model is obtained through faster training. Therefore, by adopting the reinforcement learning model optimization method, the model exploration efficiency can be improved, and model convergence can be accelerated. The method comprises the following steps: S1, acquiring a group of data from a game simulator, and inputting the data into a model; S2, the model makes a decision on simulator scene data; S3, adding various noises with different Gaussian distributions to the decision model parameters; S4, making a plurality of different decision-making behaviors through S3; S5, calculating loss for multiple different behaviors in the step S4, and carrying out averaging to update model parameters; S6, circulating S1;And S5, obtaining an optimal model until the loss is not reduced.

Description

technical field [0001] The present invention relates to the field of artificial intelligence deep reinforcement learning, in particular to a parameter noise-based reinforcement learning model optimization method and device. Background technique [0002] Reinforcement learning or Reinforcement Learning (Reinforcement Learning) has a wide range of applications and rich application prospects. The Google DeepMind team implemented a reinforcement learning model that surpasses humans in the field of Go; the industry also has extremely high application value in unmanned driving and robotic arm control. [0003] Reinforcement learning is to learn how to make corresponding behaviors according to the specific environment, so as to achieve the purpose of obtaining the maximum reward. Unlike supervised learning, reinforcement learning has no labels, but by learning the returns obtained from exploring unknown behaviors and discovering (exploitation) historical behaviors in a specific en...

AI Technical Summary

Problems solved by technology

[0007] The main purpose of the present invention is to propose a parameter noise-based reinforcement learning model optimization method and device to overcome the problems of low recognition accuracy and slow convergence in existing value-based and policy-based methods

Images