notesum.ai

Published at November 2

Task-Aware Harmony Multi-Task Decision Transformer for Offline Reinforcement Learning

cs.LG

cs.AI

Released Date: November 2, 2024

Authors: Ziqing Fan¹, Shengchao Hu¹, Yuhang Zhou¹, Li Shen¹, Ya Zhang¹, Yanfeng Wang¹, Dacheng Tao¹

Aff.: ¹Institution not specified

Arxiv: http://arxiv.org/abs/2411.01146v1

Refer to caption

Method	Meta-World 50 Tasks
#Partition	Near-optimal	Sub-optimal	Params
CARE (online)	$46.12_{\pm 1.30}$	-	1.26 M
PaCo (online)	$54.31_{\pm 1.32}$	-	3.39 M
Soft-M (online)	$53.41_{\pm 0.72}$	-	1.62 M
D2R (online)	$63.53_{\pm 1.22}$	-	1.40 M
PCGrad*	$46.34_{\pm 1.73}$	$27.04_{\pm 2.11}$	1.74 M
CAGrad*	$45.60_{\pm 1.34}$	$27.63_{\pm 1.82}$	1.74 M
MTBC	$60.39_{\pm 0.86}$	$34.53_{\pm 1.25}$	1.74 M
MTIQL	$56.21_{\pm 1.39}$	$43.28_{\pm 0.90}$	1.74 M
MTDIFF-P	$59.53_{\pm 1.12}$	$48.67_{\pm 1.32}$	5.32 M
MTDIFF-P-ONEHOT	$61.32_{\pm 0.89}$	$\textbf{48.94}_{\pm 0.95}$	5.32 M
MTDT	$20.99_{\pm 2.66}$	$20.63_{\pm 2.21}$	0.87 M
MTDT*	$65.80_{\pm 1.02}$	$42.33_{\pm 1.89}$	1.47 M
MGDT*	$66.63_{\pm 1.11}$	$42.82_{\pm 1.03}$	2.15 M
Prompt-DT	$45.68_{\pm 1.84}$	$39.76_{\pm 2.79}$	0.87 M
Prompt-DT*	$\textbf{69.33}_{\pm 0.89}$	$48.40_{\pm 0.16}$	1.47 M
HarmoDT-R (ours)	$\textbf{75.39}_{\pm 1.18}$	$\textbf{53.80}_{\pm 1.07}$	1.47 M
HarmoDT-M (ours)	$\textbf{80.33}_{\pm 0.97}$	$\textbf{57.20}_{\pm 0.73}$	1.47 M
HarmoDT-F (ours)	$\textbf{82.20}_{\pm 0.40}$	$\textbf{57.20}_{\pm 0.68}$	1.47 M