Dynamic control and optimization model of accounting information disclosure quality of listed companies based on deep reinforcement learning

Abstract

The goal of this study is to develop an intelligent dynamic control and optimization model using Deep Reinforcement Learning (DRL) technology to address the issues of low efficiency and poor adaptability of static models in environmental Accounting Information Disclosure (AID) audits of listed companies. Based on the theory of data quality and process management, the dual recording mechanism of operation and maintenance behavior log is designed, and the audit data collection and storage system is built. On this basis, Deep Q Network (DQN) is selected as the core algorithm, and a dual playback mechanism and target network are combined to construct an intelligent audit model that adapts to a high-dimensional state space, and a compound reward function including audit accuracy, disclosure quality change, and audit cost is designed. The data in the Wind database and the annual reports of 200 listed companies in Shanghai and Shenzhen A-shares from 2018 to 2022 are used in empirical studies. According to the experimental data, the average accuracy of this research model in disclosure quality identification and violation detection tasks is 88.67 and 86.50%, respectively, which is obviously superior to four mainstream algorithms, namely K-Nearest Neighbors (KNN), Bidirectional Long Short-Term Memory (Bi-LSTM), Proximal Policy Optimization (PPO), and Advantage Actor-Critic (A2C). After the model is optimized and suggested to be rectified, the average improvement rate of sample disclosure quality exceeds 30%, and the ablation experiment and sample disturbance experiment verify the effectiveness and strong anti-interference ability of the core module of the model. The study provides a brand-new technical path and practical reference for the intelligent control of AID in the capital market.