Adaptive advantage learning method and application thereof

Abstract

The application discloses a self-adaptive advantage learning method, comprising the following steps: S1, introducing a dynamic scaling factor in a target update expression of a value function, and self-adapting weighting of an advantage term; S2, when an action gap is less than or equal to a preset threshold, the dynamic scaling factor tends to 1, the advantage term is reserved to enhance the action gap; S3, when the action gap is greater than the preset threshold, the dynamic scaling factor tends to 0, the advantage term is inhibited or removed, and the target update tends to a Bellman optimal operator. Compared with the prior art, the application has the advantages that a self-adaptive advantage learning method and application thereof are provided, a theoretical optimal balance between action gap enhancement required by robustness and faster convergence speed of a value function is achieved by introducing a self-adaptive dynamic scaling factor, and a technical defect of slow convergence speed of an existing AL operator is overcome.

Images