notesum.ai

Published at October 30

Offline Behavior Distillation

cs.LG

cs.AI

Released Date: October 30, 2024

Authors: Shiye Lei¹, Sen Zhang¹, Dacheng Tao²

Aff.: ¹School of Computer Science, The University of Sydney; ²College of Computing & Data Science, Nanyang Technological University

Arxiv: http://arxiv.org/abs/2410.22728v1

Refer to caption

	Arch/Opt	Halfcheetach			Hopper			Walker2D			Average
	Arch/Opt	M-R	M	M-E	M-R	M	M-E	M-R	M	M-E	Average
Architecture	$2$ -layer	$37.1$	$35.9$	$10.9$	$29.9$	$26.2$	$33.9$	$49.2$	$41.3$	$51.1$	$35.1$ ( $\small 3.1$ )
	$3$ -layer	$38.6$	$39.7$	$19.4$	$39.0$	$28.1$	$41.5$	$63.2$	$44.1$	$55.3$	$41.0$ ( $\small 2.8$ )
	$5$ -layer	$36.1$	$37.7$	$20.0$	$37.1$	$29.1$	$36.6$	$52.0$	$36.7$	$31.6$	$35.2$ ( $\small 3.0$ )
	$6$ -layer	$32.1$	$36.0$	$17.3$	$36.9$	$29.6$	$32.8$	$47.1$	$28.2$	$25.5$	$31.7$ ( $\small 6.5$ )
	Residual	$36.9$	$36.4$	$20.0$	$38.8$	$29.8$	$40.3$	$47.5$	$35.7$	$37.1$	$35.8$ ( $\small 2.4$ )
Optim	Adam	$35.8$	$37.6$	$22.9$	$40.5$	$31.2$	$40.2$	$55.8$	$41.9$	$47.7$	$39.3$ ( $\small 1.1$ )
	AdamW	$36.8$	$37.9$	$21.4$	$40.6$	$33.3$	$41.1$	$55.4$	$44.2$	$43.2$	$39.3$ ( $\small 1.1$ )
	SGDm	$36.4$	$37.3$	$21.8$	$40.4$	$30.9$	$39.2$	$54.7$	$40.2$	$42.1$	$38.1$ ( $\small 0.1$ )