Incremental Recursive Ranking Grouping – A Decomposition Strategy for Additively and Nonadditively Separable Problems Marcin M. Komarnicki Department of Systems and Computer Networks Wroclaw University of Science and Technology Wroclaw, Poland marcin.komarnicki@pwr.edu.pl Michal W. Przewozniczek Department of Systems and Computer Networks Wroclaw University of Science and Technology Wroclaw, Poland michal.przewozniczek@pwr.edu.pl Halina Kwasnicka Department of Artifcial Intelligence Wroclaw University of Science and Technology Wroclaw, Poland halina.kwasnicka@pwr.edu.pl Krzysztof Walkowiak Department of Systems and Computer Networks Wroclaw University of Science and Technology Wroclaw, Poland krzysztof.walkowiak@pwr.edu.pl ABSTRACT Many real-world optimization problems may be classifed as Large- Scale Global Optimization (LSGO) problems. When these high- dimensional problems are continuous, it was shown efective to embed a decomposition strategy into a Cooperative Co-Evolution (CC) framework. The efectiveness of the method that decomposes a problem into subproblems and optimizes them separately may depend on the decomposition accuracy and cost. Recent decomposi- tion strategy advances focus mainly on Diferential Grouping (DG). However, when a considered problem is nonadditively separable, DG-based strategies may report some variables as interacting, al- though the interaction between them does not exist. Monotonicity checking strategies do not sufer from this disadvantage. However, they sufer from another decomposition inaccuracy ś monotonicity checking strategies may miss discovering many existing interac- tions. Therefore, Incremental Recursive Ranking Grouping (IRRG) is a new proposition that accurately decomposes both additively and nonadditively separable problems. The decomposition cost of IRRG is higher when compared with Recursive DG 3 (RDG3). Since the higher cost was a negligible part of the overall computational budget, optimization results of the considered CC frameworks were afected mainly by the decomposition accuracy. CCS CONCEPTS · Computing methodologies Artifcial intelligence; Con- tinuous space search. KEYWORDS Large-Scale Global Optimization, Problem Decomposition, Mono- tonicity Checking, Nonadditive Separability Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for proft or commercial advantage and that copies bear this notice and the full citation on the frst page. Copyrights for third-party components of this work must be honored. For all other uses, contact the owner/author(s). GECCO ’23 Companion, July 15–19, 2023, Lisbon, Portugal © 2023 Copyright held by the owner/author(s). ACM ISBN 979-8-4007-0120-7/23/07. https://doi.org/10.1145/3583133.3595846 ACM Reference Format: Marcin M. Komarnicki, Michal W. Przewozniczek, Halina Kwasnicka, and Krzysztof Walkowiak. 2023. Incremental Recursive Ranking Grouping ś A Decomposition Strategy for Additively and Nonadditively Separable Problems. In Genetic and Evolutionary Computation Conference Companion (GECCO ’23 Companion), July 15–19, 2023, Lisbon, Portugal. ACM, New York, NY, USA, 2 pages. https://doi.org/10.1145/3583133.3595846 1 INCREMENTAL RECURSIVE RANKING GROUPING Incremental Recursive Ranking Grouping (IRRG) [2] is a recursive decomposition strategy [5] that is derived from monotonicity check- ing [1]. Its typical time complexity is O(log()) . According to [1] and when a -dimensional function : Ω R is considered, two disjoint groups of variables 1 and 2 are interacting if 1 , 2 > 0, u 1 1 , u 2 2 , x * Ω such that ( x * )≤ ( x * + 1 u 1 )∧ ( x * + 2 u 2 ) > ( x * + 1 u 1 + 2 u 2 ) (1) where unit vector u j = [ 1 , ..., ] is a member of only when 1,..., = 0 . Monotonicity checking strategies may be classifed as empirical linkage learning techniques [4] because the above interaction check ensures that only existing interactions between decision variables can be discovered [2]. In comparison to other decomposition strategies, IRRG intro- duces mainly three new mechanisms. (I) Creation of two rankings r 1 = [ 1,1 , ...,  1, ] and r 2 = [ 2,1 , ...,  2, ] , where is a user- defned parameter that indicates the number of samples, to check if 1 and 2 interact. To generate samples, at frst we need values for each variable 1 that are evenly taken from the feasible set. Let us denote these values as ˜ 1, , ..., ˜ , . Then, vectors ˜ x i = [ ˜ ,1 , ..., ˜ , ] can be created. Value ˜ , is defned as ˜ , = ˜ () , , 1 * , 1 (2) where is a random permutation of {1, ...,  } for the th variable and * is the th element of a given vector x * . The ranking are gen- erated based on these vectors afterward. The th value of r 1 ( 1, ) is computed for ( ˜ x i ) . Value 2, is based on ( ˜ x i + 2 u 2 ) . Two dis- joint sets of variables 1 and 2 interact if * ∈{ 1,..., } 1, * 2, * . (II) Applying an initial optimization to fnd a high-quality solution 27