Game environment automatic decomposition method adaptive to hierarchical reinforcement learning

Abstract

The invention discloses a game environment automatic decomposition method adaptive to hierarchical reinforcement learning, and relates to two aspects, one is a weak supervision semantic segmentation technology in the aspect of computer vision, is a task decomposition module, and the other is the field of hierarchical reinforcement learning in reinforcement learning, and is a strategy training module. According to the method, the learning difficulty of reinforcement learning can be greatly reduced, so that a stronger game AI can be trained in a more complex game environment. Therefore, on one hand, the productivity of a game company in the aspect of designing the game AI can be improved, and on the other hand, the applicability of the reinforcement learning technology can be improved, so that the reinforcement learning technology can be further applied to more general fields.

Description

technical field [0001] The invention relates to a method for automatically decomposing a game environment adapted to layered reinforcement learning, which involves computer vision in machine learning and relevant field knowledge of reinforcement learning. Background technique [0002] With the increasing development of reinforcement learning, the demand for reinforcement learning to be implemented in real scenarios is daily strong. At present, reinforcement learning has a good application in training game AI, such as Go, StarCraft, Glory of Kings, QQ Speed, etc. These mainly adopt existing reinforcement learning algorithms, such as the DQN algorithm. However, reinforcement learning algorithms have great learning difficulties in complex environments, which are manifested in the slowness and instability of learning. For this reason, on the one hand, everyone is trying to invent more powerful reinforcement learning algorithms, and on the other hand, everyone is also trying to...

AI Technical Summary

Problems solved by technology

But in this case, the first is that people need manual labor. The more complex the environment, the more energy they spend on manual design. It is so that every environment requires repeated human labor. Even similar environments require repeated decomposition and design by humans, resulting in unnecessary waste of labor.

Images