Target region reduction algorithm based on improved DQN model

This paper presents a target region reduction algorithm based on an improved DQN model to address the issue of inefficient control of water jet direction when utilizing intelligent devices for fire extinguishing or mud flushing. Firstly, by leveraging the enhanced UNet network and incorporating self-similarity pooling and anti-pooling methods, the DQN model significantly improves the extraction capability of the target region in the environmental image. Secondly, the ConvLSTM network is employed as the agent of the DQN model to generate effective memory associated with the past environments and actions which contains image sequence information encompassing both temporal and spatial dimensions. Ultimately, the algorithm described in this paper achieves efficient control of the water jet direction, resulting in a 12.1% reduction in the number of water jet instances during simulation experiments compared to the optimal values obtained from four other comparative algorithms.