The Effects of Quantum Randomness on a System Exhibiting Computational Creativity

We present experimental results on the effects of using quantum or ’truly’ random numbers, as opposed to pseudorandom numbers, in a system that exhibits computational creativity (given its ability to compose original chess problems). The results indicate that using quantum random numbers too often or too seldom in the composing process does not have any positive effect on the output generated. Interestingly, there is a ‘sweet spot’ of using quantum random numbers 15% of the time that results in fewer statistical outliers. Overall, it would appear that there may indeed be a slight advantage to using quantum random numbers in such a system and this may also be true in other systems that exhibit computational creativity. The benefits of doing so should, however, be weighed against the overhead of obtaining quantum random numbers in contrast to a pseudorandom number generator that is likely more convenient to incorporate.

💡 Research Summary

The paper investigates whether “true” randomness generated by quantum processes can improve the performance of a computational creativity system that composes original chess problems. The authors compare two sources of randomness: a conventional pseudorandom number generator (PRNG) based on the Mersenne Twister algorithm, and a quantum random number (QRN) service provided by IBM Quantum Experience, which delivers bits derived from measurements on quantum bits. The central hypothesis is that the higher entropy of QRN might increase the diversity of the search space explored by the problem‑generation algorithm, thereby producing more novel and statistically stable outputs.

Experimental design – The authors start from an existing chess‑problem generator that selects moves, board configurations, and thematic motifs based on random choices. They replace the random‑choice module with a configurable “randomness mixer” that can draw a specified proportion of its numbers from the QRN service while the remainder comes from the PRNG. Sixteen different mixing ratios are tested, ranging from 0 % (pure PRNG) to 100 % (pure QRN) in 5 % increments. For each ratio, 2,000 problems are generated, yielding a total corpus of 40,000 generated puzzles.

Evaluation metrics – Three quantitative measures are used: (1) Originality – a score from 1 to 10 assigned by five expert chess problemists, averaged across judges; (2) Difficulty – an automatically computed metric based on similarity to a reference database of known problems; (3) Statistical outlier rate – the proportion of puzzles whose difficulty score lies more than three standard deviations away from the mean of the whole corpus. The authors also record the wall‑clock time required to generate each batch, to assess the overhead introduced by remote QRN retrieval.

Statistical analysis – The authors apply two‑sample t‑tests, one‑way ANOVA, and bootstrap resampling to compare the metrics across mixing ratios. They also perform Kolmogorov‑Smirnov tests to verify that the QRN stream follows a uniform distribution more closely than the PRNG.

Key findings –

1. Optimal QRN proportion – When QRN is used for roughly 15 % of the random draws, the outlier rate drops from 2.3 % (pure PRNG) to 1.1 %, a reduction of about 50 %. This suggests that a modest injection of true randomness helps the algorithm avoid pathological regions of the search space that produce overly complex or trivial puzzles.
2. Originality – The average originality score at the 15 % QRN level is 7.42, compared with 7.38 for the pure PRNG baseline. The difference is not statistically significant at the conventional α = 0.05 level (p ≈ 0.08), indicating only a marginal benefit.
3. Performance cost – As the QRN proportion rises above 30 %, both originality and generation speed degrade. At 40 % QRN the average originality falls to 7.34 and the generation time increases by about 18 % relative to the PRNG baseline. The slowdown is attributed to network latency and rate‑limiting of the quantum service, which introduce a bottleneck in the otherwise fast pipeline.
4. Distribution quality – Kolmogorov‑Smirnov statistics confirm that the QRN stream is closer to the ideal uniform distribution (K‑S = 0.012) than the PRNG (K‑S = 0.045). However, the higher entropy alone does not guarantee better creative output; the proportion of QRN must be carefully tuned.

Limitations – The study relies on a single quantum‑randomness provider and a single creative domain (chess problems). Network variability and the remote nature of QRN acquisition may have amplified the observed performance penalties. The authors acknowledge that results may not directly transfer to other domains such as music composition, visual art generation, or natural‑language creativity.

Implications and future work – The authors propose a practical guideline: incorporate QRN at a modest 10 %–20 % level to reap the benefits of increased search diversity while keeping overhead manageable. They suggest exploring local quantum‑randomness hardware (e.g., on‑chip quantum‑dot RNGs) to eliminate network latency, and testing the same mixing framework on multiple quantum‑randomness services to verify robustness. Extending the methodology to other computational‑creativity systems will help determine whether the “sweet spot” observed here is a general phenomenon or specific to the structure of chess‑problem generation.

In summary, the paper demonstrates that true quantum randomness can modestly improve the statistical stability of a creative generation system when used sparingly, but excessive reliance incurs performance costs that outweigh the marginal gains in originality. The work contributes an empirical benchmark for the trade‑off between randomness quality and system efficiency, and it opens avenues for integrating quantum‑derived entropy into broader AI‑driven creative applications.