The study proposes a novel action sequence predictor called Decision ConvFormer (DC), which is based on the MetaFormer architecture. DC uses local convolution filtering as the token mixer and effectively captures the inherent local associations of the Reinforcement Learning (RL) dataset. The study reveals that DC performs better than the Decision Transformer (DT) in various standard RL benchmarks and requires fewer resources. DC is also shown to better understand the underlying meaning in data and displays an enhanced generalization capability.
Publication date: 4 Oct 2023
Project Page: https://arxiv.org/abs/2310.03022v1
Paper: https://arxiv.org/pdf/2310.03022
