Neural and Evolutionary Computing 19 JAN, 2012

Set-based Multiobjective Fitness Landscapes: A Preliminary Study

Set-based Multiobjective Fitness Landscapes: A Preliminary Study

Fitness landscape analysis aims to understand the geometry of a given optimization problem in order to design more efficient search algorithms. However, there is a very little knowledge on the landscape of multiobjective problems. In this work, following a recent proposal by Zitzler et al. (2010), we consider multiobjective optimization as a set problem. Then, we give a general definition of set-based multiobjective fitness landscapes. An experimental set-based fitness landscape analysis is conducted on the multiobjective NK-landscapes with objective correlation. The aim is to adapt and to enhance the comprehensive design of set-based multiobjective search approaches, motivated by an a priori analysis of the corresponding set problem properties.

💡 Research Summary

The paper addresses a fundamental gap in the study of multi‑objective optimization (MOO): while fitness‑landscape analysis is well‑established for single‑objective problems, the geometry of MOO problems has received far less attention. Building on the “set‑problem” perspective introduced by Zitzler et al. (2010), the authors treat a multi‑objective solution as a set of individual solutions rather than a single point in decision space. This shift enables the definition of a set‑based multi‑objective fitness landscape (SMOFL), where the search space consists of subsets of the underlying decision space, the fitness function evaluates the quality of an entire set (using a chosen indicator such as hyper‑volume, ε‑indicator, or R2), and neighborhood operators act on the set level (e.g., replace one element, add or remove an element).

The theoretical contribution consists of three parts. First, a formal definition of the landscape is given:

  • Search space Ω = 2^X, where X is the set of all feasible solutions; practical algorithms restrict themselves to fixed‑size populations (μ‑sized subsets).
  • Set‑fitness function F : Ω → ℝ, defined by any quality indicator that is monotone with respect to Pareto dominance.
  • Neighborhood structure N(S) : for a set S, N includes all sets obtained by a single elementary operation (replace, add, or delete).

Second, the authors discuss how classical landscape measures (ruggedness, autocorrelation, plateau ratio, basin size) can be extended to the set level by aggregating over the fitness differences of neighboring sets.

Third, an experimental framework is built on the multi‑objective NK‑landscape model. The classic NK model (N = 12 binary variables, K ∈ {0,2,4,6,8,10}) is extended to M = 2 or 3 objectives, with a controllable correlation coefficient ρ ∈ {‑0.9,‑0.5,0,0.5,0.9} between objectives. This allows systematic exploration of how objective correlation and epistatic interaction (K) shape the set‑based landscape.

The empirical study proceeds as follows. For each combination of (K, ρ, M) the authors generate 30 independent instances, fix the population size μ at 50, 100, and 150, and evaluate the landscape using hyper‑volume as the primary set‑fitness indicator. They compute: (i) the distribution of hyper‑volume values across the landscape, (ii) the proportion of plateaus (neighbouring sets with identical fitness), (iii) the mean and standard deviation of fitness differences between neighbours (a proxy for ruggedness), and (iv) the average length of greedy ascent paths from random initial sets to local optima.

Key findings are:

  1. Objective correlation dominates landscape smoothness. Positive correlation (ρ > 0) yields low plateau ratios and small fitness‑difference variance, indicating a smooth landscape where high‑quality sets are reachable via small, incremental changes. Negative correlation (ρ < 0) creates many plateaus and high variance, reflecting a fragmented landscape with many local optima.

  2. Epistasis (K) increases ruggedness but its effect is mitigated at the set level. As K grows, single‑solution NK landscapes become more rugged, yet the aggregation over μ solutions smooths the set‑fitness surface, especially for larger μ.

  3. Population size influences exploration cost and landscape perception. Larger μ improves the chance of finding high‑hyper‑volume sets and reduces the number of steps needed to reach a local optimum, but the computational burden grows roughly as O(μ·N). This suggests a trade‑off and motivates adaptive population‑size strategies.

  4. Neighbourhood operator choice matters. The “replace‑one‑element” operator excels for fine‑grained exploitation in positively correlated settings, while the “add/delete” operators provide broader jumps that help escape plateaus in negatively correlated or highly epistatic instances.

These insights have direct implications for the design of set‑based MOO algorithms such as SMS‑EMOA, MOEA/D‑DE, and recent hyper‑volume‑based evolutionary strategies. For problems with strong objective conflict (negative ρ) or high epistasis, algorithms should favour larger, more diverse populations and incorporate aggressive set‑level mutations (add/delete). Conversely, when objectives are aligned, smaller, more focused populations with subtle replacements can achieve faster convergence.

The paper concludes that set‑based fitness‑landscape analysis is a viable tool for a priori algorithm design in multi‑objective optimization. Future work is outlined: extending the methodology to larger‑scale, real‑world MOO benchmarks; investigating other quality indicators (e.g., IGD, R2) within the landscape framework; exploring dynamic adaptation of μ and neighbourhood operators based on online landscape measurements; and integrating landscape‑aware mechanisms into existing MOEAs to improve robustness across diverse problem classes.