Nested constraint on binary operators?

[B-LIE]

I'm curious: how can I do nested constraints on binary operators in Julia?

I've tried:

nested_constraints=[/ => [/ => 0]],

but this gives an error message pointing to the fact that / is not a unary operator. At the same time, the documentation for nested_constraints suggests that it is possible to apply nested_constraints also to binary operators.

[MilesCranmer]

The way to do it is [(/) => [(/) => (0, 0)]]. In other words, you need to provide a tuple of integers rather than a single integer. These integers correspond to the left and right argument. This is so that you can, e.g., set a different constraint for the base of ^ than the exponent.

[B-LIE]

Thanks -- this takes me part of the way. I now get an error message when I try to do fit!:

using SymbolicRegression: SRRegressor
using MLJ: machine, fit!, predict, report, scitype, schema
#
x = range(0,2pi,length=100)
x = collect(x)
X = (x = x,)  
y = exp.(-x)
# 
model_exp2 = SRRegressor(
    batching=true,
    niterations=400,
    binary_operators=[+,-,*,/],
    nested_constraints=[(/) => [(/) => (0,0)]],
    maxsize=30,
    timeout_in_seconds=60.0 *5, 
    save_to_file=false,
)
# 
mach = machine(model_exp2, X, y)
fit!(mach)

leads to:

{
	"name": "MethodError",
	"message": "MethodError: no method matching isless(::Int64, ::Tuple{Int64, Int64})\n\nClosest candidates are:\n  isless(!Matched::Missing, ::Any)\n   @ Base missing.jl:87\n  isless(::Any, !Matched::Missing)\n   @ Base missing.jl:88\n  isless(!Matched::Tuple{}, ::Tuple)\n   @ Base tuple.jl:555\n  ...\n",
	"stack": "MethodError: no method matching isless(::Int64, ::Tuple{Int64, Int64})\n\nClosest candidates are:\n  isless(!Matched::Missing, ::Any)\n   @ Base missing.jl:87\n  isless(::Any, !Matched::Missing)\n   @ Base missing.jl:88\n  isless(!Matched::Tuple{}, ::Tuple)\n   @ Base tuple.jl:555\n  ...\n\n\nStacktrace:\n  [1] <(x::Int64, y::Tuple{Int64, Int64})\n    @ Base .\\operators.jl:352\n  [2] <=(x::Int64, y::Tuple{Int64, Int64})\n    @ Base .\\operators.jl:401\n  [3] >=(x::Tuple{Int64, Int64}, y::Int64)\n    @ Base .\\operators.jl:425\n  [4] SymbolicRegression.CoreModule.OptionsStructModule.Options(; binary_operators::Vector{Function}, unary_operators::Vector{Function}, constraints::Nothing, elementwise_loss::Nothing, loss_function::Nothing, tournament_selection_n::Int64, tournament_selection_p::Float64, topn::Int64, complexity_of_operators::Nothing, complexity_of_constants::Nothing, complexity_of_variables::Nothing, parsimony::Float64, dimensional_constraint_penalty::Nothing, dimensionless_constants_only::Bool, alpha::Float64, maxsize::Int64, maxdepth::Nothing, turbo::Bool, bumper::Bool, migration::Bool, hof_migration::Bool, should_simplify::Nothing, should_optimize_constants::Bool, output_file::Nothing, node_type::Type, populations::Int64, perturbation_factor::Float64, annealing::Bool, batching::Bool, batch_size::Int64, mutation_weights::SymbolicRegression.CoreModule.MutationWeightsModule.MutationWeights, crossover_probability::Float64, warmup_maxsize_by::Float64, use_frequency::Bool, use_frequency_in_tournament::Bool, adaptive_parsimony_scaling::Float64, population_size::Int64, ncycles_per_iteration::Int64, fraction_replaced::Float64, fraction_replaced_hof::Float64, verbosity::Nothing, print_precision::Int64, save_to_file::Bool, probability_negate_constant::Float64, seed::Nothing, bin_constraints::Nothing, una_constraints::Nothing, progress::Nothing, terminal_width::Nothing, optimizer_algorithm::Optim.BFGS{LineSearches.InitialStatic{Float64}, LineSearches.BackTracking{Float64, Int64}, Nothing, Nothing, Optim.Flat}, optimizer_nrestarts::Int64, optimizer_probability::Float64, optimizer_iterations::Nothing, optimizer_f_calls_limit::Nothing, optimizer_options::Nothing, use_recorder::Bool, recorder_file::String, early_stop_condition::Nothing, timeout_in_seconds::Float64, max_evals::Nothing, skip_mutation_failures::Bool, nested_constraints::Vector{Pair{typeof(/), Vector{Pair{typeof(/), Tuple{Int64, Int64}}}}}, deterministic::Bool, define_helper_functions::Bool, deprecated_return_state::Nothing, fast_cycle::Bool, npopulations::Nothing, npop::Nothing, kws::@Kwargs{})\n    @ SymbolicRegression.CoreModule.OptionsModule C:\\Users\\Bernt\\.julia\\packages\\SymbolicRegression\\9q4ZC\\src\\Options.jl:598\n  [5] get_options(m::SRRegressor{DynamicQuantities.SymbolicDimensions{DynamicQuantities.FixedRational{Int32, 25200}}, DataType})\n    @ SymbolicRegression.MLJInterfaceModule C:\\Users\\Bernt\\.julia\\packages\\SymbolicRegression\\9q4ZC\\src\\MLJInterface.jl:66\n  [6] update(m::SRRegressor{DynamicQuantities.SymbolicDimensions{DynamicQuantities.FixedRational{Int32, 25200}}, DataType}, verbosity::Int64, old_fitresult::Nothing, old_cache::Nothing, X::@NamedTuple{x::Vector{Float64}}, y::Vector{Float64}, w::Nothing)\n    @ SymbolicRegression.MLJInterfaceModule C:\\Users\\Bernt\\.julia\\packages\\SymbolicRegression\\9q4ZC\\src\\MLJInterface.jl:124\n  [7] fit\n    @ C:\\Users\\Bernt\\.julia\\packages\\SymbolicRegression\\9q4ZC\\src\\MLJInterface.jl:119 [inlined]\n  [8] fit(m::SRRegressor{DynamicQuantities.SymbolicDimensions{DynamicQuantities.FixedRational{Int32, 25200}}, DataType}, verbosity::Int64, X::@NamedTuple{x::Vector{Float64}}, y::Vector{Float64})\n    @ SymbolicRegression.MLJInterfaceModule C:\\Users\\Bernt\\.julia\\packages\\SymbolicRegression\\9q4ZC\\src\\MLJInterface.jl:119\n  [9] fit_only!(mach::MLJBase.Machine{SRRegressor{DynamicQuantities.SymbolicDimensions{DynamicQuantities.FixedRational{Int32, 25200}}, DataType}, SRRegressor{DynamicQuantities.SymbolicDimensions{DynamicQuantities.FixedRational{Int32, 25200}}, DataType}, true}; rows::Nothing, verbosity::Int64, force::Bool, composite::Nothing)\n    @ MLJBase C:\\Users\\Bernt\\.julia\\packages\\MLJBase\\qETMX\\src\\machines.jl:692\n [10] fit_only!\n    @ C:\\Users\\Bernt\\.julia\\packages\\MLJBase\\qETMX\\src\\machines.jl:617 [inlined]\n [11] #fit!#63\n    @ C:\\Users\\Bernt\\.julia\\packages\\MLJBase\\qETMX\\src\\machines.jl:789 [inlined]\n [12] fit!(mach::MLJBase.Machine{SRRegressor{DynamicQuantities.SymbolicDimensions{DynamicQuantities.FixedRational{Int32, 25200}}, DataType}, SRRegressor{DynamicQuantities.SymbolicDimensions{DynamicQuantities.FixedRational{Int32, 25200}}, DataType}, true})\n    @ MLJBase C:\\Users\\Bernt\\.julia\\packages\\MLJBase\\qETMX\\src\\machines.jl:786\n [13] top-level scope\n    @ c:\\Users\\Bernt\\OneDrive\\...\\Notebooks\\jl_notebook_cell_df34fa98e69747e1a8f8a730347b8e2f_X64sZmlsZQ==.jl:2"
}

[KushilaJayamanne]

The API documentation says that "For binary operators, both arguments are treated the same way, and the max of each argument is constrained." So, I'm not sure if we can apply a different constraint to each argument, but it is possible to apply the same constraint to both arguments.

E.g., nested_constraints = [(*)=> [(+)=> 0]] specifies that a multiplication x*y cannot have a + in either x or y. So, expressions like x1 * (x2 + x3) and (x1 + x2) * x3 are not allowed.

[MilesCranmer]

Sorry! I mixed up two APIs (constraints and nested_constraints). Indeed the right answer is [(/) => [(/) => 0]]