An optimisation framework for audit decision-making based on deep convolutional neural networks and reinforcement learning

In contemporary auditing and risk management, transaction complexity and high-dimensional audit data challenge accurate risk assessment and efficient resource allocation.Traditional methods relying on manual expertise or static heuristics fail in large-scale, multimodal, dynamic audit scenarios.To solve these, this paper proposes RLMCN-Net, a framework combining multimodal convolutional neural networks and deep reinforcement learning for audit risk assessment and action optimisation.Its risk identification module estimates audit targets' underlying risk from heterogeneous data; the reinforcement learning module determines audit actions considering costs, resource constraints, and expected returns.Risk labels, audit actions, and decision outcomes are modelled separately.Experiments on two benchmark datasets show RLMCN-Net outperforms traditional baselines in risk identification accuracy, audit return, and resource efficiency; ablation studies verify its robustness and generalisation.These results indicate RLMCN-Net effectively supports risk-oriented target screening and dynamic resource allocation, advancing intelligent auditing systems.