Intelligent agent automatic decision-making method based on reinforcement learning

Abstract

The invention discloses an intelligent agent automatic decision-making method based on reinforcement learning. Determining an environment state and an action space of the intelligent agent; constructing a target network for determining a first selection probability of the selectable action and an evaluation network for determining a post-effect reward value for implementing the first selection probability action, determining a current environment state and a target environment state of the intelligent agent, actions of each step of the intelligent agent are obtained through the target networkand the evaluation network; and forming a trajectory strategy, storing the trajectory strategy into an experience sample pool, generating a new trajectory strategy according to an existing trajectorystrategy in the experience sample pool to expand the experience sample pool, and performing parameter updating on the evaluation network and the target network by adopting samples in the experience sample pool according to a preset updating period. By adopting the method, the neural network training performance can be improved, so that the automatic decision accuracy of the intelligent agent is improved.

Description

technical field [0001] The invention belongs to the technical field of machine learning, and more specifically relates to an automatic decision-making method for an intelligent body based on reinforcement learning. Background technique [0002] Reinforcement Learning (RL) originally originated from psychology and is used to imitate the learning mode of intelligent creatures. It is a special type of model-free machine learning that takes the state of the environment (State) as input and aims to adapt to the environment (Environment). method. figure 1 It is a schematic diagram of the reinforcement learning process. Such as figure 1 As shown, the core idea of ​​reinforcement learning is to use the feedback signal (Reward) obtained from the environment to optimize a series of policies (Policy) through continuous interaction with the environment and continuous trial and error (Explorer). [0003] Reinforcement learning has been involved and applied to many fields, such as auto...

AI Technical Summary

Problems solved by technology

[0005] In the existing method, since only the relevant trajectory strategies that achieve the specified goal state are stored in the experience sample pool, the relevant trajectory strategies that fail to achieve the desired goal in the specified number of explorations and specified time steps are discarded, which may cause the constructed experience The available data in the sample pool is small, that is to say, the number of effective samples is small, the rewards achieved by the strategy in the trajectory are sparse, and a lot of resources are wasted, including the cost of manually designing reward functions, coding costs, hardware facilities, etc.

Images