Skip to content
BrainlessLab BrainlessLab BrainlessLab

Evolve a structured CTRNN

Use this path when you want to search a fixed compartmental continuous-time recurrent neural network (CTRNN). The example uses a small SepCMA budget so that you can inspect the complete workflow.

It does not evolve topology, body structure, or ports. It also does not establish that the selected model is optimal.

Build the plan

Save this Julia file as make_structured_ctrnn_plan.jl:

using BrainlessLab


composition = CompositionSpec(
    :structured_ctrnn_wall,
    :compartmental_structured,
    :wall;
    n_nodes=60,
)


training = EvaluationTarget(
    :wall_development,
    composition,
    EvaluationSpec(
        blocks=1,
        trials_per_block=2,
        horizon=240,
        warmup=40,
        construction_scope=:trial,
        reset=:full,
        root_seed=0x1101,
        aggregate=:mean,
    ),
)


heldout = EvaluationTarget(
    :wall_heldout,
    composition,
    EvaluationSpec(
        blocks=1,
        trials_per_block=2,
        horizon=240,
        warmup=40,
        construction_scope=:trial,
        reset=:full,
        root_seed=0x2201,
        aggregate=:mean,
    ),
)


run = BrainlessLab.Evolution.RunConfig(;
    strategy=:sepcma,
    iterations=2,
    search_seed=0x2a,
    measure=:normalized_score,
    direction=:maximise,
    initialisation=BrainlessLab.Evolution.NormalInitialisation(
        centre=:zero,
        scale=0.25,
    ),
    options=(population=4, reducer=:minimum,),
)


plan = EvolutionPlan(
    :structured_ctrnn_search,
    (training,);
    run=run,
    heldout_targets=(heldout,),
)


write_plan("plans/structured_ctrnn_search.toml", plan)

The node key :compartmental_structured selects the fixed StructuredCompartmental design. Use :compartmental_dense when the declared question requires the fixed DenseCompartmental design.

Other registered nodes can use EvolutionPlan when their NodeSpec declares a reviewed Evolution.NodeDesignSpec.

The normal initialisation is explicit. search_seed controls the initial population and later search decisions. The target root seeds control evaluation trials.

This workflow uses Wall because its 200-scored-tick minimum fits the existing 240-tick horizon and 40-tick warm-up. Wall is a near-ceiling floor-check, so this smoke-budget search demonstrates the operation contract rather than a meaningful optimisation claim.

Validate and run

Run the Julia file once to write the plan:

Terminal window
julia --project=. make_structured_ctrnn_plan.jl

Then validate the generated TOML without simulation:

Terminal window
julia --project=. bin/brainlesslab.jl check \
  plans/structured_ctrnn_search.toml

Run the plan into a record root:

Terminal window
julia -t auto --project=. bin/brainlesslab.jl run \
  plans/structured_ctrnn_search.toml --root records

The command prints the record directory. Keep that exact path for the next tutorials.

Check the result

A complete run contains DONE, not INCOMPLETE or FAILED. Start with:

  • request.toml for the plan that you submitted;
  • resolved.toml for the complete resolved search;
  • seeds.csv for search and evaluation seed records;
  • the operation tables listed by record.toml;
  • the generated report for a compact inspection view.

SepCMA records one stable model role named selected. This role identifies development output. It is not a confirmed benchmark result.

Next, inspect and resume the evolution record.

Source: src/evolution/Evolution.jl, src/operations/Evolution.jl, src/records/Records.jl.