SymbolicRegression.jl

Appearance

Migrating from v1 to v2 ​

SymbolicRegression.jl v2 restructures the internals around a composable plugin interface, first-class mutation types, and customizable crossover operations. Almost all v1 code still runs without modification: the breaking surface is behavioral rather than syntactic. This page lists what to check when upgrading from v1.13.

Behavioral changes ​

Your code will run unchanged, but search results will differ:

  • Adaptive mutation weights are now on by default. Mutation weights adapt over the course of the search via AdaptiveMutationWeightsPlugin. To recover static v1 weights, choose the default plugin set explicitly:

    julia
    Options(;
    default_plugins=(SimulatedAnnealingPlugin(; alpha=3.17), AdaptiveParsimonyPlugin())
)

  • batching now defaults to :auto. Large datasets are evaluated in batches automatically, with batch_size chosen for you when not set. Restore v1 behavior with batching=false, batch_size=50.

  • crossover_probability now defaults to 0.20 (was 0.0259).

  • Simplification recomputes cost. After a tree is simplified, its cost is recomputed from scratch, so PopMember.cost can differ from v1 for an equivalent expression.

  • Constant-optimization restarts can now escape zero-valued constants, so optimization trajectories (and therefore results) differ from v1.

All other defaults are unchanged from v1.13.

Renames (deprecated shims in place) ​

These still work but emit deprecation warnings:

  • eval_options= is now eval_context= in evaluation entry points.

  • use_recorder/recorder_file are now use_tracing/tracing_file; recorder output is streamed as JSONL traces.

  • PopMember.score is now PopMember.cost.

  • camelCase keyword arguments (mutationWeights, useFrequency, shouldOptimizeConstants, ...) are deprecated in favor of snake_case.

Removed ​

  • ParametricExpression is removed. Use TemplateExpression with optimizable parameters instead.

  • Options, SearchState, and TemplateExpressionSpec gained type parameters, which changes their concrete type arity. SearchState also replaces all_running_search_statistics with plugin_states.

  • The internal mutation helpers delete_random_op! and _random_op were generalized to n-ary operators. This only affects code that defines custom mutations against internals.

v1 keyword arguments that still work ​

These are converted automatically to the new plugin/mutation configuration, and need no changes:

  • mutation_weights (including a plain vector via the deprecated mutationWeights spelling) is converted to weighted first-class mutations.

  • annealing / alpha inject a SimulatedAnnealingPlugin.

  • use_frequency / use_frequency_in_tournament configure the AdaptiveParsimonyPlugin.

  • perturbation_factor / probability_negate_constant are applied to the default ConstantMutation.

New APIs ​

  • Plugins: Options(; plugins=(MyPlugin(),)) composes search behaviors. See the Plugins page for the hook set and a worked example.

  • First-class mutations: pass mutations=[MutateConstant() => 1.0, ...] or subtype AbstractMutation; see Customization.

  • Custom crossovers: subtype AbstractCrossover and pass crossovers=[MyCrossover() => 1.0, ...].

  • MLJ-free interface: machine, fit!, predict, and report work without loading MLJ.