Featured image of post lambda debug

lambda debug

在实验中的某个已不可考证的环节,我们曾猜测提高 lambda 可以提升效果,因此 alpha=0.1 和 0.05 的 lambda 被从5提升到了20。然而,尽管这个优化没有被采纳到 V7 ,在之后的实验中我们既没有把 alpha=0.3 和 0.5 的 lambda 提升到 20 ,也没有把 alpha=0.1 和 0.05 的 lambda 恢复到5。这个事实对于之前得出的一些实验结论提出了挑战。

lambda debug

在实验中的某个已不可考证的环节,我们曾猜测提高 lambda 可以提升效果,因此 alpha=0.1 和 0.05 的 lambda 被从5提升到了20。然而,尽管这个优化没有被采纳到 V7 ,在之后的实验中我们既没有把 alpha=0.3 和 0.5 的 lambda 提升到 20 ,也没有把 alpha=0.1 和 0.05 的 lambda 恢复到5。这个事实对于之前得出的一些实验结论提出了挑战。

除此之外,我们意外发现在之前的所有实验中,服务器端的 IFFI 都没有被启用。

考虑到这两点,我们认为重新实验是有必要的。

Alpha = 0.05

Algorithm: OneshotOursV7 (DRCL with ETF Anchors and Lambda Annealing) + Simple Server Aggregation

RoundAlgorithmAccuracyModel Variance (Mean)G-Protos Std
0OursV7+Simple0.31110.0005001.00230
10OursV7+Simple0.48500.0014340.94234
20OursV7+Simple0.56800.0020260.88015
30OursV7+Simple0.62120.0025080.81882
40OursV7+Simple0.64840.0029240.76189
50OursV7+Simple0.67770.0032930.70978

Algorithm: OneshotOursV6 (DRCL and Lambda Annealing) + Simple Server Aggregation

RoundAlgorithmAccuracyModel Variance (Mean)G-Protos Std
0OursV6+Simple0.30070.0005101.00220
10OursV6+Simple0.50930.0014310.94247
20OursV6+Simple0.60190.0020270.87989
30OursV6+Simple0.63650.0025080.81893
40OursV6+Simple0.65640.0029230.76179
50OursV6+Simple0.68170.0032960.70938

Algorithm: OneshotOurs (V4) + Simple Server Aggregation

RoundAlgorithmAccuracyModel Variance (Mean)G-Protos Std
0OursV4+Simple0.28290.0005091.00606
10OursV4+Simple0.49730.0014301.00622
20OursV4+Simple0.57840.0020121.00636
30OursV4+Simple0.62380.0024771.00648
40OursV4+Simple0.63840.0028801.00661
50OursV4+Simple0.66970.0032341.00672

Algorithm: OneshotOursV6 (DRCL and Lambda Annealing) + Advanced IFFI Server Aggregation

RoundAlgorithmAccuracyModel Variance (Mean)G-Protos Std
0OursV6+Advanced0.25200.0005101.00220
10OursV6+Advanced0.51210.0014310.94247
20OursV6+Advanced0.61410.0020270.87989
30OursV6+Advanced0.67360.0025080.81893
40OursV6+Advanced0.69750.0029230.76179
50OursV6+Advanced0.72110.0032960.70938

Algorithm: OneshotOurs (V4) + Advanced IFFI Server Aggregation

RoundAlgorithmAccuracyModel Variance (Mean)G-Protos Std
0OursV4+Advanced0.24450.0005091.00606
10OursV4+Advanced0.49490.0014301.00622
20OursV4+Advanced0.59160.0020121.00636
30OursV4+Advanced0.64470.0024771.00648
40OursV4+Advanced0.65900.0028801.00661
50OursV4+Advanced0.69780.0032341.00672

Alpha = 0.1

Algorithm: OneshotOursV7 (DRCL with ETF Anchors and Lambda Annealing) + Simple Server Aggregation

RoundAlgorithmAccuracyModel Variance (Mean)G-Protos Std
0OursV7+Simple0.31580.0005561.00131
10OursV7+Simple0.56200.0016030.92845
20OursV7+Simple0.65960.0022830.85353
30OursV7+Simple0.71610.0028300.78167
40OursV7+Simple0.74350.0033050.71536
50OursV7+Simple0.76860.0037300.65452

Algorithm: OneshotOursV6 (DRCL and Lambda Annealing) + Simple Server Aggregation

RoundAlgorithmAccuracyModel Variance (Mean)G-Protos Std
0OursV6+Simple0.30800.0005581.00132
10OursV6+Simple0.54470.0016070.92906
20OursV6+Simple0.64530.0022910.85352
30OursV6+Simple0.70150.0028160.78187
40OursV6+Simple0.73430.0033150.71540
50OursV6+Simple0.75950.0037420.65502

Algorithm: OneshotOurs (V4) + Simple Server Aggregation

RoundAlgorithmAccuracyModel Variance (Mean)G-Protos Std
0OursV4+Simple0.31320.0005541.00604
10OursV4+Simple0.54650.0016131.00621
20OursV4+Simple0.65340.0022781.00637
30OursV4+Simple0.70240.0028161.00651
40OursV4+Simple0.73210.0032831.00664
50OursV4+Simple0.76100.0036971.00676

Algorithm: OneshotOursV7 (DRCL with ETF Anchors and Lambda Annealing) + Advanced IFFI Server Aggregation

RoundAlgorithmAccuracyModel Variance (Mean)G-Protos Std
0OursV7+Advanced0.28790.0005541.00604
10OursV7+Advanced0.55750.0016131.00621
20OursV7+Advanced0.66570.0022781.00637
30OursV7+Advanced0.70670.0028161.00651
40OursV7+Advanced0.74470.0032831.00664
50OursV7+Advanced0.77200.0037031.00676

Algorithm: OneshotOursV6 (DRCL and Lambda Annealing) + Advanced IFFI Server Aggregation

RoundAlgorithmAccuracyModel Variance (Mean)G-Protos Std
0OursV6+Advanced0.29770.0005581.00132
10OursV6+Advanced0.55870.0016130.92906
20OursV6+Advanced0.66140.0022780.85352
30OursV6+Advanced0.71620.0028180.78187
40OursV6+Advanced0.75360.0032850.71540
50OursV6+Advanced0.77890.0037030.65502
保留所有权。
comments powered by Disqus