notesum.ai

Published at December 6

cs.RO

cs.HC

Released Date: December 6, 2024

Authors: Charles Dietzel¹, Patrick J. Martin²

Aff.: ¹Commonwealth Center for Advanced Manufacturing, Disputanta, VA, USA; ²Department of Computer Science, University of Richmond, VA, USA

	IBC-1	IBC-2	Searched Values
Hyperparameter
Training iterations	100000	100000
Batch size	512	512
Sequence length	3	3
Network size (width x depth)	512x4	512x4	512x4, 256x8, 128x16
Activation function	ReLU	ReLU
Learning rate	1e-3	1e-3	1e-3, 5e-4, 1e-4
Learning rate decay	0.99	0.99
Learning rate decay steps	100	100
Gradient penalty	Final step only	Final step only
Gradient margin	1	1
Training counter-examples	8	4	8, 4
Langevin iterations	100	100
Langevin learning rate init.	0.1	0.1
Langevin learning rate final	1e-5	1e-5
Langevin polynomial decay power	2	2
Langevin delta action clip	0.1	0.1
Selected by	Soft DTW-GI	GDTW, Soft GDTW, DTW-GI