ASS-DQN: A deep reinforcement learning approach for path planning in multi-machine cooperative operations

With the advancement of smart agriculture and uncrewed operations, deep reinforcement learning (DRL) has been increasingly applied to path planning and task allocation in multi-agent collaborative operations. Conventional Deep Q-Networks (DQN), however, face challenges in multi-machine agricultural scheduling, including overestimation of Q-values, redundant experience samples, and limited exploration efficiency, leading to unstable decision-making and suboptimal task allocation in dynamic environments. To address these issues, this study proposes the Alternating Maximum Q Deep Reinforcement Learning algorithm (ASS-DQN), which integrates three key mechanisms: an Alternating max-softmax target strategy (AltMaxQ) to mitigate Q-value overestimation, a Minimum Uncertainty Selection (SMU) mechanism to prioritize high-value experiences and enhance training efficiency, and a Suboptimal Solution Group (SSG) mechanism to improve exploration and global decision-making.A multi-machine collaborative simulation environment based on cotton harvesting was developed to evaluate ASS-DQN. Experimental results show that, in large-scale fault-free operations, ASS-DQN outperforms Greedy allocation, Ant Colony Optimization (ACO), and Genetic Algorithm (GA), achieving reductions exceeding 20% in decision rewards, energy consumption, and system latency. Compared with Q-learning, Proximal Policy Optimization(PPO) and DQN, ASS-DQN reduced operation time by 13.33%, 4.20%, and 6.60%, respectively, and total energy consumption by 12.26%, 8.14%, and 5.63%, respectively, and system latency by 9.76%, 15.14%, and 7.03%, respectively. Under fault scenarios, ASS-DQN also surpasses DQN, Dueling DQN, PPO, and D3QN across these metrics. These findings indicate that ASS-DQN effectively enhances multi-machine collaborative task allocation, improves operational efficiency, and reduces resource consumption in dynamic agricultural environments.