Skip to content
neuroflow0.2.31 · docs

← neuroflow

/neuroflow:brain-optimize

Run a parameter search or fit a computational brain model to experimental data.

/brain-optimize covers two modes: systematic parameter sweeps to map model behaviour, and data fitting to match experimental recordings.

When to use it

  • You want to explore how model outputs change across a range of parameter values
  • You want to fit model parameters to experimental data (spike rates, LFP, EEG, BOLD)
  • You need a structured optimisation pipeline with a defined cost function

Two modes

Claude asks which you want:

Systematically explore a parameter space to map model behaviour.

Claude asks: - Which parameters to sweep and their ranges (e.g. synaptic weight 0.1–2.0 mS, time constant 5–50 ms) - What metric to record per run (mean firing rate, burst index, oscillation frequency, correlation) - How many parameter combinations — grid, random, or Latin hypercube sampling - Run locally or via HPC/cluster

Output: a sweep spec run by sweep_run.py, which iterates over the parameter grid, runs the model for each combination, and saves per-run metrics to results.csv and the best configurations to summary.json. A smoke configuration runs first; the sweep is refused until it passes.

Optimise model parameters to match target experimental data.

Claude asks: - What experimental data to fit to (file path or description) - What features to match (firing rate, ISI distribution, LFP power spectrum, coherence, ERP amplitude) - What optimisation algorithm (grid search, Nelder-Mead, differential evolution, Bayesian optimisation, DEAP/evolutionary) - What is the cost function - How many iterations / evaluations are acceptable

Output: an optimisation script with a cost function, parameter bounds, and optimiser loop.

Steps

  1. Write optimize-plan.md — parameters being searched, target features, algorithm, cost function, convergence criteria
  2. Implement the optimisation script, or the sweep spec
  3. Run a minimal test (2–3 parameter combinations) to confirm the pipeline works end-to-end — for sweeps, the smoke gate of sweep_run.py
  4. After a full run: summarise best-fit parameters, cost value, and whether convergence was reached — computed from the results files, not read off a plot. A grid sweep returns the best grid point, and sweep_run.py warns when it sits on the edge of the grid

Plans and result summaries go to .neuroflow/brain-optimize/. Scripts and raw results go to your models/optimize/ folder. Long sweeps run in the background or as cluster jobs and are tracked in .neuroflow/brain-optimize/runs.md.

Files read and written

Direction Files
Reads .neuroflow/project_config.md, .neuroflow/flow.md, .neuroflow/brain-optimize/flow.md, .neuroflow/brain-build/flow.md
Writes .neuroflow/brain-optimize/, .neuroflow/brain-optimize/flow.md, .neuroflow/reasoning/brain-optimize.jsonl, .neuroflow/sessions/YYYY-MM-DD.md, models/optimize/ (scripts and results)
  • /brain-build — assemble the model before optimising
  • /brain-run — run the model with best-fit parameters after optimisation