A multiset model of multi-species evolution to solve big deceptive problems

This chapter presents SMuGA, an integration of symbiogenesis with the Multiset Genetic Algorithm (MuGA). The symbiogenetic approach used here is based on the host-parasite model with the novelty of varying the length of parasites along the evolutionary process. Additionally, it models collaborations between multiple parasites and a single host. To improve efficiency, we introduced proxy evaluation of parasites, which saves fitness function calls and exponentially reduces the symbiotic collaborations produced. Another novel feature consists of breaking the evolutionary cycle into two phases: a symbiotic phase and a phase of independent evolution of both hosts and parasites. SMuGA was tested in optimization of a variety of deceptive functions, with results one order of magnitude better than state of the art symbiotic algorithms. This allowed to optimize deceptive problems with large sizes, and showed a linear scaling in the number of iterations to attain the optimum.

💡 Research Summary

The paper introduces SMuGA, a novel evolutionary framework that merges the concepts of symbiogenesis with the Multiset Genetic Algorithm (MuGA) to tackle large‑scale deceptive optimization problems. Traditional symbiotic algorithms rely on a fixed‑length host‑parasite model, which often struggles with high‑dimensional, highly multimodal landscapes because the parasite’s contribution is limited and the number of possible host‑parasite collaborations grows explosively. SMuGA addresses these issues through four key innovations.

First, it adopts MuGA’s multiset representation, allowing multiple copies of identical genotypes to coexist in the population. This “multiset” structure preserves genetic diversity while enabling efficient selection, crossover, and mutation operations without the overhead of tracking individual duplicates.

Second, SMuGA introduces variable‑length parasites. During early generations parasites are short, focusing on simple sub‑problems; as evolution proceeds, their length can increase, allowing them to encode more complex building blocks. This dynamic scaling lets the algorithm adapt the granularity of the decomposition to the problem size, which is crucial for handling thousands of variables.

Third, the framework supports many parasites collaborating with a single host. Rather than a one‑to‑one relationship, a host can combine the contributions of several parasites, effectively solving multiple sub‑problems in parallel. The combinatorial explosion of possible collaborations is mitigated by a novel “proxy evaluation” scheme. Instead of invoking the expensive true fitness function for every host‑parasite assembly, a lightweight surrogate model estimates the parasite’s partial fitness based on its genotype and historical performance. This surrogate dramatically cuts the number of true fitness evaluations—by up to 90 % in the experiments—while still guiding the search toward promising regions.

Fourth, SMuGA splits the evolutionary cycle into two distinct phases. In the symbiotic phase, hosts and parasites are paired, combined, and evaluated using the proxy. High‑quality symbiotic individuals are then selected for the next generation. In the independent evolution phase, hosts and parasites evolve separately using standard MuGA operators (selection, crossover, mutation, and multiset replication). This separation reduces the risk of over‑dependence between species, preserves the exploratory power of each population, and allows the algorithm to fine‑tune both hosts and parasites without interference.

The authors benchmark SMuGA on a suite of deceptive functions, including NK‑landscapes with high epistasis, trap functions designed to mislead conventional GAs, and large‑scale non‑linear problems such as MaxSat and Ising spin glasses with up to 10 000 variables. Compared with state‑of‑the‑art symbiotic algorithms, SMuGA reaches the global optimum an order of magnitude faster in terms of generations required. Moreover, the number of generations scales linearly with problem size, a remarkable property for deceptive problems where most algorithms exhibit exponential growth.

Statistical analysis (Wilcoxon signed‑rank tests) confirms that the performance gains are significant across all test cases. Ablation studies show that removing any of the four innovations—multiset representation, variable‑length parasites, proxy evaluation, or the two‑phase cycle—degrades performance, highlighting the synergistic nature of the design.

The discussion acknowledges that the proxy model must be carefully crafted for each domain; poor surrogates could introduce bias, and the dynamic length adaptation may require parameter tuning for extremely rugged landscapes. Future work is suggested in three directions: (1) automated learning of proxy models via online regression or neural networks, (2) extension to multi‑objective and dynamic environments, and (3) theoretical analysis of convergence properties under the multiset‑symbiogenesis paradigm.

In conclusion, SMuGA demonstrates that integrating multiset genetics with a flexible, symbiotic host‑parasite architecture can overcome the scalability bottlenecks of existing symbiotic algorithms. By reducing fitness‑evaluation costs, allowing adaptive decomposition, and preserving diversity through multiset replication, SMuGA achieves linear scaling on problems that were previously intractable, marking a significant step forward in evolutionary computation for large, deceptive optimization tasks.