eqhunt

pip install eqhunt

Symbolic regression by genetic programming. C++ engine, Python bindings via nanobind.

Give it a table of (inputs, target) pairs; it returns a human-readable formula that approximates the relationship. No neural network, no black box — just an algebraic expression you can read, paste into a calculator, or hand-tune.

import eqhunt

X = [[1, 1], [2, 3], [4, 5], [7, 2], [9, 9]]
y = [2, 5, 9, 9, 18]

model = eqhunt.fit(X, y)
print(model.formula)        # e.g.  f(x,y) = (x+y)
print(model.error)          # e.g   0.0
print(model.predict([6, 7])) # -> 13.0

Install

pip install eqhunt

Prebuilt wheels are published for Linux, macOS and Windows on common Python versions. If pip falls back to building from source you'll need a C++17 compiler.

Two ways to use it

Ultra-simple

import eqhunt

model = eqhunt.fit(X, y, generations=5000)
print(model.formula)
model.predict([1, 2])         # single row
model.predict([[1, 2], [3, 4]])  # batch

fit() accepts any Config field as a keyword argument:

eqhunt.fit(X, y, pop=800, trig_penalty=2.0, bloat_penalty=0.3)

Fully configurable

import eqhunt

cfg = eqhunt.Config()
cfg.pop               = 800
cfg.gen               = 50000
cfg.tournament_size   = 5
cfg.initial_depth     = 5
cfg.bloat_penalty     = 0.3
cfg.trig_penalty      = 1.5
cfg.accepted_error    = 0.01

# Re-weight individual operators (higher = more likely to appear)
cfg.op_weights.sin = 1.0      # boost sine
cfg.op_weights.cos = 1.0
cfg.op_weights.exp = 0.0      # disable exp entirely
cfg.pi_prob = 0.10            # 'pi' more frequent in terminals

model = eqhunt.Model(cfg).fit(X, y)
print(model.formula)

You can also train from a CSV file (one row per sample, last column = target, lines starting with # are comments):

eqhunt.Model().fit_csv("nivel_embase.csv")

Operators available

Category	Operators
Arithmetic	+ - * / -x
Powers	sqrt **
Conditional	if(cond, then, else) (cond > 0)
Trig	sin cos tan
Exp / log	exp log
Constants	numeric literals, pi

Trigonometric, log and exp nodes have low default weights so they only appear after enough mutation pressure — useful for cyclic / physical data, ignored otherwise. Adjust via Config.op_weights.

How error and validity are handled

• Per-sample error is |prediction - target|; total error is the sum.
• Invalid evaluations (/0, sqrt(<0), log(<=0), exp(huge)) get a soft per-sample penalty rather than killing the whole formula — a single out-of-domain sample no longer disqualifies an otherwise good candidate. If more than 25% of samples fail, the formula is rejected.

Stopping early

Config.accepted_error stops the search as soon as total error drops below the threshold. You can also call model.stop() from another thread (or a signal handler) to ask the loop to wrap up after the current generation.

Saving and reloading a formula

A trained model is just a string — you can persist it, ship it, paste it, diff it. To reuse a formula in a new process without retraining, parse it back into a Model:

import eqhunt

# train and save
m = eqhunt.fit(X, y)
print(m.formula)          # e.g.  f(x,y) = ((x*x) - (y*y))
m.save("model.txt")       # one-liner persisted

# later, in a fresh process — no training needed
m2 = eqhunt.Model.load_file("model.txt")
m2.predict([6, 7])        # -13.0
m2.predict([[1, 2], [3, 4]])

You can also go through strings directly:

formula_str = m.formula                   # or any equivalent expression
m3 = eqhunt.Model.from_formula(formula_str)
m3.predict([12, 5])

Or mutate an existing model in place:

m.load_formula("(x*x + y*y)")             # replaces the current tree

Accepted syntax: anything the engine itself emits via get_formula() — arithmetic (+ - * / **), unary minus, sqrt sin cos tan log exp if, variables x y z w v u x6 x7 …, numeric literals (int / float / 1e5), and pi. Both the bare expression ("(x+y)") and the full prefixed form ("f(x,y) = (x+y)") are accepted; the parser strips everything up to and including the first =. Parse errors raise RuntimeError.

The number of input variables is inferred from the highest variable index in the formula, so m2.num_vars is set correctly without needing to know it in advance.

Config reference

Field	Default	Meaning
pop	400	Population size
gen	15000	Max generations
tournament_size	4	Tournament selection pool
crossover_prob	0.7	Crossover probability per pair
mutation_prob	0.25	Mutation probability per offspring
initial_depth	4	Depth used to seed the initial population
mutation_depth	3	Depth for mutation-generated subtrees
const_min/max	-9, 9	Range for random numeric terminals
pi_prob	0.01	Probability a terminal is pi
bloat_penalty	0.1	Per-node penalty (favours smaller trees)
trig_penalty	0.5	Extra penalty per sin/cos/tan/log/exp node
immigrant_rate	0.05	Fraction of population replaced by random each gen
weak_parent_rate	0.2	Prob. 2nd parent is random (not tournament)
accepted_error	0.5	Stop training once total error < this value
verbose	False*	Print best-so-far per improvement
simplify	True	Run algebraic simplification on the final tree
simplify_interval	500	Periodically simplify top-N members during training
simplify_top_n	10	How many to simplify periodically

*C++ default is True; the Python fit() helper defaults to False.

Building from source

git clone https://github.com/sha0coder/eqhunt
cd eqhunt
pip install -e .
pytest

Requires Python 3.8+, a C++17 compiler, CMake 3.15+.

License

MIT.