Concetti Chiave
This work proves near-tight runtime guarantees for the SEMO, global SEMO, SMS-EMOA, and NSGA-III algorithms on four classic multi-objective benchmark problems, showing that these algorithms cope much better with many objectives than previously thought.
Sintesi
The paper presents a mathematical runtime analysis of several multi-objective evolutionary algorithms (MOEAs) on four classic benchmark problems: OneMinMax (mOMM), CountingOnesCountingZeros (mCOCZ), LeadingOnesTrailingZeros (mLOTZ), and OneJumpZeroJump (mOJZJk).
The key highlights and insights are:
The authors prove near-tight runtime guarantees for the SEMO, global SEMO, SMS-EMOA, and NSGA-III algorithms on these benchmarks, showing that their performance scales linearly with the size of the Pareto front, in contrast to previous quadratic bounds.
The results suggest that MOEAs cope much better with many-objective problems than previously thought, and that the performance loss observed in experiments is more likely due to the increasing size of the Pareto front rather than inherent algorithmic difficulties.
The authors identify three key properties of the algorithms that enable the transfer of their results to other MOEAs, and demonstrate this by extending the bounds to the SEMO, SMS-EMOA, and NSGA-III.
The results for mLOTZ do not improve over previous bounds, but the authors note that their analysis applies to arbitrary numbers of objectives, unlike the previous constant-objective results.
For mOJZJk, the authors provide bounds that are only applicable when the number of blocks m' is at least 2, and refer to previous results for the case m' = 1.
Overall, the paper presents a significant advancement in the theoretical understanding of many-objective evolutionary optimization, with near-tight runtime guarantees for several prominent algorithms on classic benchmark problems.
Statistiche
The size of the Pareto front for the benchmarks are:
mOMM: SOMM^m = (n/m' + 1)^m'
mCOCZ: SCOCZ^m = (n/2m' + 1)^m'
mLOTZ: SLOTZ^m = (n/m' + 1)^m'
mOJZJk: SOJZJ^m,k = (n/m' - 2k + 3)^m'
Citazioni
"Together with the parallel and independent work on the NSGA-III [25], these are the first that tight runtime guarantees for these MOEAs for general numbers of objectives, and they improve significantly over the previous results with their quadratic dependence on the Pareto front size."
"We are optimistic that our methods can also be applied to other MOEAs. We discuss some sufficient conditions for transferring the obtained bounds. We argue that they are fulfilled by SMS-EMOA and NSGA-III, but we believe that our arguments can also be applied to variants of the NSGA-II that do not suffer from the problems detected in [29], e.g., the NSGA-II with the tie-breaker proposed in [15], and to the (μ + 1) SIBEA [5]."