http://arxiv.org/abs/2211.11461
Symbolic Regression (SR) algorithms learn analytic expressions which both accurately fit data and, unlike traditional machine-learning approaches, are highly interpretable. Conventional SR suffers from two fundamental issues which we address in this work. First, since the number of possible equations grows exponentially with complexity, typical SR methods search the space stochastically and hence do not necessarily find the best function. In many cases, the target problems of SR are sufficiently simple that a brute-force approach is not only feasible, but desirable. Second, the criteria used to select the equation which optimally balances accuracy with simplicity have been variable and poorly motivated. To address these issues we introduce a new method for SR — Exhaustive Symbolic Regression (ESR) — which systematically and efficiently considers all possible equations and is therefore guaranteed to find not only the true optimum but also a complete function ranking. Utilising the minimum description length principle, we introduce a principled method for combining these preferences into a single objective statistic. To illustrate the power of ESR we apply it to a catalogue of cosmic chronometers and the Pantheon+ sample of supernovae to learn the Hubble rate as a function of redshift, finding $\sim$40 functions (out of 5.2 million considered) that fit the data more economically than the Friedmann equation. These low-redshift data therefore do not necessarily prefer a $\Lambda$CDM expansion history, and traditional SR algorithms that return only the Pareto-front, even if they found this successfully, would not locate $\Lambda$CDM. We make our code and full equation sets publicly available.
D. Bartlett, H. Desmond and P. Ferreira
Tue, 22 Nov 22
2/83
Comments: 14 pages, 6 figures, 2 tables. Submitted to IEEE Transactions on Pattern Analysis and Machine Intelligence