Application of Transformer for Encoding States in Reinforcement Learning
D. A. Kozlov · Optoelectronics Instrumentation and Data Processing · 2024
Abstract The application of the transformer architecture for state encoding in reinforcement learning algorithms is studied. A novel approach that integrates transformers with existing methods such as SAC (soft actor-critic) to improve their performance and generalization ability is presented. Experimental results show that the approach can improve learning in complex 3D locomotion acquisition tasks.