TABLE 4.

For each combination of a deep learning model, a learning rate schedule, and an optimization method in the unloaded system test presented in Sections A to C, the summary statistics of the average estimation error across the five runs.

Deep learning model	Learning rate schedule	Optimization method	Average estimation error
GoogLeNet	fixed learning rate	SGD	0.093±0.012
	fixed learning rate	AdaGrad	0.093±0.008
	exponential decay	RMSprop	0.990±0.282
	exponential decay	SGD	0.897±0.274
	exponential decay	AdaGrad	0.632±0.197
	step decay	RMSprop	0.364±0.025
	step decay	SGD	0.540±0.160
	step decay	AdaGrad	0.552±0.196
GRU	fixed learning rate	RMSprop	0.275±0.101
	fixed learning rate	SGD	0.793±0.326
	fixed learning rate	AdaGrad	0.536±0.226
	exponential decay	RMSprop	0.284±0.046
	exponential decay	SGD	0.695±0.530
	exponential decay	AdaGrad	0.631±0.360
	step decay	Adam	0.333±0.050
	step decay	RMSprop	0.384±0.133
	step decay	SGD	0.568±0.217
	step decay	AdaGrad	0.304±0.169