QMIX REINFORCEMENT LEARNING ALGORITHM-BASED SHIP WELDING SPOTS COLLABORATIVE WELDING METHOD USING MULTIPLE MANIPULATORS

摘要

A qmix reinforcement learning algorithm-based ship welding spots collaborative welding method using multiple manipulators. The method comprises the following steps: a) building a reinforcement learning environment, and setting a welding area and an operation area in the environment; b) determining state values and action values of the manipulators; c) setting reward values according to the state values, the action values, and tasks of collaborative welding and collision avoidance; d) calculating a local action value function of each manipulator according to the state values and action values and by means of a recurrent neural network, and performing an action selection process; e) obtaining an overall action value function of all manipulators from the action value functions by means of a super-network set with non-negative weights; and f) constructing a loss function from the reward values of step c) and the overall action value function network of step e), calculating and updating the weights of the neural network according to a back-propagation algorithm, and repeating the training process. The method does not depend on any system model, is simple and effective, and can implement tasks of collaborative welding of welding spots in obstacle environments.