Roulette-wheel selection via stochastic acceptance

Roulette-wheel selection is a frequently used method in genetic and evolutionary algorithms or in modeling of complex networks. Existing routines select one of N individuals using search algorithms of O(N) or O(log(N)) complexity. We present a simple roulette-wheel selection algorithm, which typically has O(1) complexity and is based on stochastic acceptance instead of searching. We also discuss a hybrid version, which might be suitable for highly heterogeneous weight distributions, found, for example, in some models of complex networks. With minor modifications, the algorithm might also be used for sampling with fitness cut-off at a certain value or for sampling without replacement.

💡 Research Summary

**
The paper introduces a novel implementation of roulette‑wheel selection, a cornerstone operation in genetic algorithms and preferential‑attachment network models, that achieves constant‑time (O(1)) average complexity by replacing the traditional search‑based approach with a “stochastic acceptance” scheme. In the classic method, a cumulative fitness array is built and a random number r∈(0, Σw_i) is generated; locating the sector containing r requires either a linear scan (O(N)) or a binary search (O(log N)). The proposed algorithm proceeds in two simple steps: (1) pick an individual uniformly at random from the population of size N, (2) accept this pick with probability w_i / w_max, where w_max is the current maximal fitness. If the pick is rejected, the process repeats from step 1.

Mathematically, the probability that individual i is finally selected after an arbitrary number of attempts is
p′_i = (w_i / (N·w_max)) · (1 + q + q² + …)
with q = 1 − ⟨w⟩/w_max, where ⟨w⟩ = Σw_i / N is the average fitness. The geometric series converges for 0 < q < 1, yielding p′_i = w_i / Σw_i, exactly the same distribution as the standard roulette‑wheel. The expected number of attempts per selection is τ = w_max / ⟨w⟩. In most practical scenarios fitness values are bounded (w_i < B) and the average fitness does not vanish as N grows, so τ remains a small constant and the overall algorithm runs in O(1) time.

Empirical tests were performed on populations where fitnesses are uniformly distributed in (0, 1). For population sizes ranging from 10² to 10⁴ the authors measured the CPU time per selection for three implementations: linear scan, binary search, and stochastic acceptance. As expected, the linear scan showed O(N) scaling, the binary search O(log N), while stochastic acceptance exhibited essentially flat time, confirming the O(1) claim. The measured τ was close to 2, matching the theoretical prediction τ ≈ w_max/⟨w⟩ ≈ 2 for the chosen distribution. Slight increases for the largest N were attributed to cache‑miss penalties, not algorithmic inefficiency.

The paper also discusses extensions. When the fitness distribution is highly skewed (e.g., one weight w₁ ≫ others), a hybrid scheme can be used: select the dominant individual directly with probability w₁/Σw_i, and apply stochastic acceptance to the remaining N‑1 individuals. This reduces τ dramatically. For sampling without replacement, the selected individual’s fitness can be set to zero; if the removed individual was the current maximum, w_max must be recomputed, but the overall complexity stays O(1). If a known upper bound B exists, the acceptance probability can be fixed to w_i/B, sacrificing a bit of efficiency while still guaranteeing constant‑time performance.

Finally, the authors note that the method can be adapted to dynamic settings where fitness values evolve over time, by periodically updating w_max or using a constant A > w_max for acceptance. They argue that because the algorithm is extremely simple, requires only two random numbers per attempt, and avoids any data‑structure dependent searches, it is well suited for high‑performance evolutionary computation and for network growth models that rely on preferential attachment. The presented results show a substantial CPU speed‑up (often an order of magnitude or more) over traditional implementations for population sizes typical in genetic algorithms (10²–10⁴). The paper concludes that stochastic acceptance provides a practical, theoretically sound, and easily extensible alternative to existing roulette‑wheel selection techniques.