Efficient three-dimensional boxing optimization method based on deep reinforcement learning

By adopting the priority experience playback mechanism and SAC algorithm in deep reinforcement learning, the network convergence and overfitting problems caused by high discretency and sparse rewards in three-dimensional packing problems are solved, and more efficient learning and performance improvement are achieved.

CN119721889APending Publication Date: 2025-03-28ZHEJIANG UNIV OF TECH +1
0 Cites 2 Cited by

Patent Information

Application Number
CN202411805350.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-12-09
Publication Date
2025-03-28

AI Technical Summary

Technical Problem

In deep reinforcement learning, the three-dimensional packing problem causes the network to converge too slowly due to high discreteness and sparse rewards, low learning efficiency, and prone to overfitting problems.

Method used

The priority experience playback mechanism is used in combination with the SAC algorithm, and the random and uniform experience playback is replaced by the compound priority experience playback mechanism, which improves the effectiveness of experience, accelerates the convergence of the network and stabilizes the training process.

Benefits of technology

It significantly improves the learning ability and performance of online three-dimensional packaging, alleviates the overfitting problem, avoids local optimization, and significantly accelerates the convergence speed of the network.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The invention belongs to the field of logistics article boxing technology and reinforcement learning, and discloses an efficient three-dimensional boxing optimization method based on deep reinforcement learning. Comprising the steps that a three-dimensional boxing problem is described as a Markov decision process, and a reward mechanism is set as follows: immediate rewards are obtained every time one article is loaded, and final rewards are obtained according to the total loading rate of containers after all the articles are loaded; the method comprises the steps that an experience pool is created, a track generated by interaction of an intelligent agent and the environment is obtained, the track comprises the state of the current moment, the action of the current moment, the immediate reward of the current moment, the state of the next moment and the TD error of the current moment, the track is stored in the experience pool, and the composite priority of the track is calculated according to the immediate reward and the TD error; and training an Actor-Critic network model according to the composite priority and applying the Actor-Critic network model. According to the invention, the performance and robustness of the strategy are improved.
Need to check novelty before this filing date? Find Prior Art