Task analysis-based few-sample multi-agent reinforcement learning generalization method

Abstract

The invention discloses a few-sample multi-agent reinforcement learning generalization method based on task analysis. Comprising the following steps: processing text / image input through a task analysis model; generating a structured sub-task allocation and semantic embedding vector by a text scheme by utilizing a large language model; the image scheme generates a distribution scheme and an image feature vector through a multi-modal model; a shared parameter sequential network is adopted to encode an agent track, and environment dynamic features are extracted through dimension reduction aggregation; splicing the subtask feature representation, the environment feature and the observation state into a joint feature vector; and calculating the local action value of each agent based on the joint features, dynamically allocating value weights through a super network, carrying out weighted fusion, and jointly optimizing network parameters by using time sequence differential loss. According to the method, a deep coupling mechanism of task target analysis and environment perception is innovatively introduced, the performance is remarkably improved in complex multi-task scene verification, and an efficient solution is provided for open environments such as multi-robot collaboration and unmanned aerial vehicle clusters.

Images