What Is the Flower Pollination Algorithm?

The flower pollination algorithm is a population-based metaheuristic for global optimization that models the reproductive behavior of flowering plants as a search process. Introduced by Xin-She Yang in 2012, it treats each candidate solution as a flower carrying pollen and alternates between two movement operators: a long-range step representing cross-pollination carried by insects and birds, and a short-range step representing self-pollination by wind or gravity. Because the two operators differ in scale rather than in kind, the algorithm has an unusually small parameter set for its class, with a single switch probability governing the balance between them.

It belongs to the family of nature-inspired metaheuristics that includes particle swarm optimization, differential evolution, and cuckoo search, and like them it makes no assumption that the objective function is differentiable, continuous, or unimodal. The original formulation presented at the 2012 Unconventional Computation and Natural Computation conference states four idealized rules that reduce the biology to a workable update scheme.

Global Pollination and Lévy Flights

Global pollination represents biotic cross-pollination, in which a pollinator carries pollen between distant plants. The corresponding update moves a solution toward the best solution found so far, scaled by a step drawn from a Lévy distribution rather than a Gaussian one. A Lévy flight produces mostly small steps punctuated by occasional very large ones, which matches observed foraging patterns in insects and gives the search a heavy tail that helps it escape local optima. In practice the step is generated by the Mantegna algorithm with a stability index near 1.5, and the resulting displacement is applied along the direction connecting the current solution to the incumbent best.

Local Pollination and Flower Constancy

Local pollination represents abiotic and self-pollination, and is implemented as a move along the difference vector between two randomly chosen solutions from the population, scaled by a uniform random number. The operator refines solutions within the region the population already occupies, so it supplies exploitation while the Lévy step supplies exploration. Which operator fires on a given iteration is decided by a switch probability, and Yang's benchmark experiments found values near 0.8 to be effective for most problems, meaning local moves dominate. A later study on flower pollination algorithm parameter tuning examined how population size and switch probability interact across benchmark suites and found the algorithm relatively forgiving of parameter choice compared with several competing metaheuristics. The fourth of Yang's rules, flower constancy, treats the probability that two flowers exchange pollen as proportional to their similarity, and it is what the difference-vector form encodes.

Variants and Extensions

The basic algorithm was written for unconstrained continuous single-objective problems, and most subsequent work has extended it past those limits. Yang's multi-objective flower pollination formulation uses weighted-sum scalarization with randomized weights to trace a Pareto front, and applies the result to structural design problems including disc brake and beam design. Binary and discrete versions replace the continuous update with a transfer function or a permutation-based operator so that the method can address feature selection, knapsack, and scheduling problems. Hybrids pair the pollination operators with local search, with the simplex method, or with the update rules of other swarm algorithms, and elite opposition-based variants generate mirrored candidate solutions to improve early coverage of the search space. As with other metaheuristics, no single variant dominates, and reported performance depends heavily on the benchmark set and the tuning effort spent on each competitor.

Applications

The flower pollination algorithm has been applied in a range of engineering problems, including:

• Antenna array synthesis and sidelobe level reduction
• Economic dispatch and unit commitment in power systems
• Photovoltaic model parameter extraction and maximum power point tracking
• Feature selection and clustering in machine learning pipelines
• Structural and mechanical design optimization
• Wireless sensor network node placement and routing