[ANN] Arborist.jl — Generic Genetic Programming for Julia

Package Announcements
1

I’m pleased to announce the initial release of Arborist.jl, a generic genetic programming framework for Julia following the Problem/Algorithm/Solve pattern.

Why now?

The Julia ecosystem has lacked a general-purpose GP framework since Wallace.jl died at Julia 0.3 in 2015. Existing packages (Metaheuristics.jl, Evolutionary.jl) target numerical optimization with vector genomes. Arborist.jl targets genetic programming — evolving trees, programs, and neural topologies — with explicit support for the LLM-as-mutation-operator pattern that FunSearch and AlphaEvolve have brought to mainstream attention.

What does it do?

Symbolic regression in two lines:

using Arborist, DynamicExpressions

evaluator = SymbolicRegressionEvaluator(
    x -> x^4 + x^3 + x^2 + x, domain=(-1f0, 1f0), points=20
)
result = solve(
    GPProblem(evaluator, TreeGenome{Float32}; seed=42),
    GeneticProgramming(pop_size=100, generations=200)
)

Five genome types cover different problem classes:

  • TreeGenome — DynamicExpressions.jl-backed, 8x faster, for symbolic regression
  • ExprGenome — Julia AST compilation via @eval, for general program synthesis
  • GraphGenome — NEAT-style neural topology with structural mutation
  • AntGenome — side-effectful agent control programs
  • ADFGenome — Koza-style Automatically Defined Functions on top of TreeGenome

NEAT-style neural topology evolution:

using Arborist
reset_innovation_counter!()
X = Float64[0 0 1 1; 0 1 0 1]; y = Float64[0 1 1 0]
ops = neat_defaults()
result = solve(
    GPProblem(GraphEvaluator(X, reshape(y,1,4)), GraphGenome; seed=42),
    GeneticProgramming(pop_size=150, generations=150,
                       mutation_ops=ops.mutation_ops,
                       crossover_ops=ops.crossover_ops,
                       speciation=ThresholdSpeciation(threshold=3.0))
)

LLM mutation operator (Anthropic, OpenAI, or local Ollama) — the FunSearch/AlphaEvolve pattern as a composable operator within a standard evolutionary loop. See the documentation for details.

Benchmarks

A representative slice of the benchmark suite (ARBORIST_RUN_BENCHMARKS=true,
~27 min wall time). Each gate is verified across 5 independent seeds.

Problem Genome Gate Passing
Koza-1 / Koza-2 / Koza-3 TreeGenome fitness < 0.1 3/5 each
Nguyen-1..6, -8..10 TreeGenome fitness < 0.01 3/5
XOR GraphGenome (NEAT) fitness < 0.01 4/5
UCI Iris (one-vs-rest) TreeGenome test acc ≥ 90% 4/5
Cart-pole GraphGenome (NEAT) ≥ 195 steps mean 4/5
Two-spirals (NSGA-II) GraphGenome best-front error < 1.0; HV > 0 —

See the README
for the full set covering 25+ problems across symbolic regression, Boolean
synthesis, classification, control tasks, modularity, time series, and
multi-objective formulations.

Related packages

Arborist.jl complements rather than replaces existing packages:

Try it

using Pkg; Pkg.add("Arborist")

A paper describing Arborist.jl with the Koza and NEAT benchmarks is in preparation. FunSearch/AlphaEvolve comparison benchmarks are planned for the next release.

4 Likes