Patent attributes
The present invention provides a multi-agent cooperation decision-making and training method, including the following steps: S1: encoding, by an encoder, local observations obtained by agents by using a multi-layer perceptron or a convolutional neural network as feature vectors in a receptive field; S2: calculating, by a graph convolution layer, relationship strength between the agents by using a relationship unit of a multi-headed attention mechanism, integrating, by a relationship convolution kernel of the relationship unit, the feature vectors in the receptive field into new feature vectors, and iterating the graph convolution layer for multiple times to obtain a relationship description of the multi-headed attention mechanism in a larger receptive field and at a higher order; S3: splicing the feature vectors in the receptive field and the new feature vectors integrated by the graph convolution layer, sending the spliced vectors to a value network, wherein the value network selects and performs an action decision with the highest future feedback expectation; and S4: storing a local observation set and related sets of the agents in a buffer region, collecting samples in the buffer region for training, and optimizing and rewriting a loss function.
