Method for improving accuracy of reliable action selection in intelligent agent control

Abstract

The invention discloses a method for improving the accuracy of selecting reliable actions in agent control, which comprises the following steps of: 1, interacting an agent deployed with a behavior strategy network, a probabilistic neural network and an evaluation scoring network in advance with a real environment according to a preset target task to be completed to collect real environment data, simulating real environment dynamics through probabilistic neural network learning from the collected real environment data to obtain a plurality of dynamic models; 2, learning a plurality of estimations of an evaluation scoring function of the evaluation scoring network by the intelligent agent based on a plurality of kinetic models; step 3, the intelligent agent optimizes the strategy of the behavior strategy network by using the average value of k minimum estimations in the plurality of estimations of the obtained evaluation scoring function; and 4, performing behavior selection by adopting an optimized strategy in intelligent agent control. The sample efficiency of a model reinforcement learning method for the intelligent agent can be improved, so that the learning strategy performance is improved, and the accuracy of selecting reliable actions in control is improved.

Description

technical field [0001] The invention relates to the field of intelligent body control, in particular to a method for improving the accuracy of selecting reliable actions in intelligent body control. Background technique [0002] Reinforcement learning has achieved great success in decision-making tasks, from playing video games to controlling robots in simulators. However, many of these results are achieved by model-free reinforcement learning methods and often require a large number of samples, which greatly hinders the application of model-free reinforcement learning methods to real-world tasks. [0003] In contrast, model-based RL methods build a model of the environment and generate fictional interactions with higher sample efficiency than model-free RL methods. Therefore, for agent learning, model-based reinforcement learning methods are more promising approaches for real-world tasks, such as robot control, industrial control, etc. The sample efficiency of modeled rei...

AI Technical Summary

Problems solved by technology

[0006] The purpose of the present invention is to provide a method for improving the accuracy of selecting reliable actions in the control of the agent, which can improve the sample efficiency of the model-based reinforcement learning method for the agent, and then solve the problem of the model-based learning method of the agent. Low efficiency, resulting in low performance of the learned strategy, resulting in poor accuracy of selecting reliable actions in control

Images