Thirteen synthetic regression tasks are registered in
Program.cs. Inputs are sampled independently
and uniformly in the order shown below. Physical constants are variable inputs;
h denotes reduced Planck's constant (ℏ).
Fifteen further AI Feynman equations are registered as a development set (developmentSets
in the same file) and are used only for pilots that choose budgets and settings; they are
not part of the evaluation.
With --noise s, training and validation targets receive Gaussian noise with standard
deviation s times the clean target's (fixed noise seed, drawn per row in row order); the test
partition is scored against the clean and the noisy targets. The study uses s = 0.1 for a
noisy subset of conditions.
| Runner identifier / descriptor | Inputs in descriptor order | Domains in that order | Target expression |
|---|---|---|---|
| I.6.20 / Feynman2 | sigma, theta | [1,3], [1,3] | exp(-theta²/(2 sigma²))/(sqrt(2 pi) sigma) |
| I.9.18 / Feynman5 | m1, m2, G, x1, x2, y1, y2, z1, z2 | [1,2]³, [3,4], [1,2], [3,4], [1,2], [3,4], [1,2] | G m1 m2 / ((x2-x1)²+(y2-y1)²+(z2-z1)²) |
| I.15.3.x / Feynman17 | x, u, c, t | [5,10], [1,2], [3,20], [1,2] | (x-u t)/sqrt(1-u²/c²) |
| I.30.5 / Feynman31 | lambd, d, n | [1,2], [2,5], [1,5] | asin(lambd/(n d)) |
| I.32.17 / Feynman33 | epsilon, c, Ef, r, omega, omega_0 | [1,2]⁵, [3,5] | (epsilon c Ef²/2)(8 pi r²/3) omega⁴/(omega²-omega_0²)² |
| I.41.16 / Feynman44 | omega, T, h, kb, c | [1,5]⁵ | h omega³/(pi² c² (exp(h omega/(kb T))-1)) |
| I.48.20 / Feynman50 | m, v, c | [1,5], [1,2], [3,20] | m c²/sqrt(1-v²/c²) |
| II.6.15a / Feynman56 | epsilon, p_d, r, x, y, z | [1,3]⁶ | 3 p_d z sqrt(x²+y²)/(4 pi epsilon r⁵) |
| II.11.27 / Feynman64 | n, alpha, epsilon, Ef | [0,1]², [1,2]² | n alpha epsilon Ef/(1-n alpha/3) |
| II.11.28 / Feynman65 | n, alpha | [0,1]² | 1+n alpha/(1-n alpha/3) |
| II.35.21 / Feynman81 | n_rho, mom, B, kb, T | [1,5]⁵ | n_rho mom tanh(mom B/(kb T)) |
| III.9.52 / Feynman90 | p_d, Ef, t, h, omega, omega_0 | [1,3]⁴, [1,5]² | p_d Ef t/h sinc²((omega-omega_0)t/2) |
| III.10.19 / Feynman91 | mom, Bx, By, Bz | [1,5]⁴ | mom sqrt(Bx²+By²+Bz²) |
sinc(z) = sin(z)/z, with sinc(0) = 1.
Definitions follow AI Feynman and the domains in Haider's 2021 preprint, with these explicit choices:
- I.6.20 uses the two-input Gaussian with denominator
sqrt(2 pi) sigma. Haider's preprint printssqrt(2 pi sigma). - I.30.5 uses
lambdin [1,2] so allasinarguments are valid. - III.9.52 includes the squared sinc denominator from AI Feynman and its continuous limit at equal frequencies.
The runner fixes the data seed at 42 and reuses the same noiseless observations across methods and repetitions. Current defaults are 8,000 training, 2,000 validation and 10,000 test rows.
GP fits on training data; HPO uses validation NMSE. Test data score the selected expression and fresh-seed runs of the nominated configuration. Fresh GP seeds measure search variability on the same data.
Run the equation, domain and partition checks from the repository root:
dotnet run --project HyperOp.Checks -c Release