Table 5.

Key constraints, assumptions, recommendations, and considerations for all SP variants.

Key constraints and assumptions for all secretary problem variants
A	The total number of neural network models within the NAS search space is known
B	The total number of neural network models within the NAS search space is at least 20 (n ≥ 20)
C	Neural network models are selected at random to be interrogated and ranked; Every neural network model within the NAS search space has an equal chance at being selected
D	Each neural network model interrogated is relatively ranked from best to worst against only previously interrogated neural network models
E	Every neural network model within the NAS search space can be uniquely ranked; No ties exist

Recommendations and considerations by secretary problem variant
Secretary problem variant	Recommend optimal policy minimum (${\tau }_{\pi }$ or r/n)	Computational resource savings (times better)	Considerations
CSP	0.368	2.7x	Interrogating 36.8% of the NAS search space returns the highest probability of success in discovering the best performing neural network model; Interrogating 36.8% of the NAS search space may not be possible for certain total population values of n; Rounding up to the nearest population unit integer may be required
GEP	0.15	6.7x	Interrogating 15% of the NAS search space gives an 80% probability of success in discovering a neural network model of Rank 10 or better; As the probability of success in discovering a neural network model of the best Rank with a certain Rank range grows, the required exploration of the NAS search space shrinks resulting in a Rank increase of the best Rank discovered
CBP	0.04	25x	Interrogating at least 4% of the NAS search space returns a better Rank on average with the “call back” feature as compared to the average Rank selected using the CSP rules; Interrogating at least 10% of the NAS search space returns a better Rank on average than the GEP optimal policy of 15%