Table 1 Algorithm characteristics and mean ranking scores of Task 1 submissions

Team	Aggregation method					lr schedule	Client selection	Score
	DS	PD	LO	LI	Combination
FLSTAR	✓		✓		⊙	Constant	6 largest	2.75
Sanctuary	✓	✓	✓		⊙	Polynomial	Alternating: all; drop slow clients	3.05
RoFL	✓	✓			⊙ + server optimizer	Step	All	3.35
gauravsingh	✓			✓	⊕	Constant	6 random	3.67
rigg	✓		✓	✓	⊕ (weighted)	Constant	Randomly drop large clients	4.65
HT-TUAS	✓	✓			⊕	Constant	4 random	4.69
Flair	✓				Multiple gradient descent with contraint	Constant	All	5.85

Algorithm characteristics include the aggregation method, learning rate (lr) schedule, and client selection. Algorithms are listed in the order of ranking score contained in the Score column, with the best on top. See the methods section for how the ranking score is calculated. A common pattern for aggregation methods is to compute multiple normalized weight terms (DS Dataset size, PD (inverse) Parameter distance, LO Potential for local optimization, LI Local improvement) and combine them either through arithmetic mean (⊕) or multiplicative averaging (⊙). The weight term abbreviations were introduced here as categories summarizing the main idea behind the weight terms, but the implementation details in the teams’ algorithms differed slightly, as described in the methods section. Only one team chose a completely different aggregation approach (Flair). Selectively sampling clients was used by five teams to improve the convergence speed.

Quick links

Search