A new family of AI algorithms can reconstruct how the universe is changing over time with dramatically greater precision than previous methods, according to a study posted to the preprint server arXiv earlier this year.
The technique, called GAME, short for Genetic Algorithms with Marginalised Ensembles, delivered a 95% improvement in accuracy when calculating derivatives, the measurements that capture how quickly cosmic quantities are changing rather than just what their current values are. For overall function reconstruction, it showed a 20% improvement. The gap between those two numbers tells the story of why this matters.
Genetic algorithms have been a tool in cosmology for some time. Modeled loosely on natural selection, they let researchers reconstruct mathematical functions directly from observational data without forcing the results to fit a predetermined theoretical model. That makes them valuable for independent checks on the standard cosmological model, the framework that accounts for the universe's accelerating expansion, dark energy, and galaxy formation.
But these algorithms have a persistent weakness. They tend to produce reliable-looking overall fits to observed data while becoming unreliable when asked to compute derivatives of those fits. Derivatives are crucial because they describe rates of change, how fast the universe is expanding, how quickly dark energy's influence is shifting, how structures are growing. Getting the overall curve right while getting its slope wrong is a problem that undermines precisely the kind of precision testing that cosmologists need.
The GAME approach addresses this by running not one algorithm but a whole ensemble of them simultaneously. Each algorithm tackles the same problem slightly differently, producing its own candidate solution. GAME then applies ensemble averaging, assigning each solution a weight based on how well it fits the data and how mathematically smooth the resulting function is. The final output is a weighted combination of all the candidates rather than a single best-guess answer.
The result is a system that is less likely to get trapped in what researchers describe as deceptive local solutions, regions where an algorithm converges on something that looks correct but is actually misleading. By pooling multiple independent runs, GAME dilutes the influence of any single algorithm's blind spots.
The timing is relevant. Cosmology is currently wrestling with several unresolved tensions in its standard model. Measurements of how fast the universe is expanding, known as the Hubble constant, differ depending on the method used to make them. Recent large-scale surveys have produced results that strain some predictions. Better tools for reconstructing cosmic behavior directly from data, without model assumptions, give researchers a sharper instrument for probing those discrepancies.
The GAME paper has not yet completed peer review, having been posted to arXiv. But if the accuracy gains hold up, the approach could become a standard tool for analyzing data from current and upcoming sky surveys that are mapping the universe in unprecedented detail.
