Create Experiment for Experiment Manager

Plans and creates experiments in MATLAB Experiment Manager by analyzing user code or problem descriptions, determining the experiment type, and generating the appropriate experiment function, hyperparameter table, and experiment object.

When to Use

• User has a MATLAB script/function and wants to sweep parameters or hyperparameters
• User asks to create an experiment or compare configurations
• User describes a problem and wants to explore parameter space
• User wants to track outputs across different parameter combinations

When NOT to Use

• User is debugging an existing experiment or training failure
• User wants to run, resume, or view results of an existing experiment
• User already has a working Experiment Manager experiment set up
• User wants to edit experiment settings (hyperparameters, strategy, description) — direct them to make changes in the Experiment Manager app. However, editing the experiment function code (.m file) is allowed.
• User wants Bayesian optimization or random sampling strategy — direct them to configure from the Experiment Manager app

---

Autonomy Model

This skill follows a guidance + autonomy pattern. Decision points are tagged:

Tag	Meaning
[user]	Always pause and ask. Non-negotiable.
[auto]	Agent decides silently. Escalates on failure.
[depends]	Resolved by clarity of signals — auto if clear, ask if ambiguous.

After analysis, tell the user how you intend to work in one plain sentence. Always add: "Let me know if you'd prefer more or less control."

---

Phase 1: Analysis

Read the user's code or problem description and classify the experiment type.

Type Classification

Classify into one of three types using API-level and semantic signals.

IMPORTANT: If semantic signals indicate training (even without explicit DL API calls), classify as training — NOT general purpose.

Type	Signals	Template class
General purpose	No DL functions, no training indicators. Optimization, simulation, data analysis.	'experiments.internal.experimentTemplates.MATLABFunction'
Built-in training	Uses trainnet/trainNetwork/trainingOptions, standard layers, datastores. OR: file name contains "train", user mentions "classification"/"transfer learning".	'experiments.internal.experimentTemplates.BlankTrainnetTraining'
Custom training (custom loop)	Uses dlnetwork/dlfeval/dlgradient/dlarray, manual gradients, GANs. OR: training loop with loss computation, user mentions "fine-tune"/"detector".	'experiments.internal.experimentTemplates.CustomTraining'

Classification Priority

1. Training loop structure (epoch iteration, loss tracking) → Custom training
2. Sets up data/layers/options without looping → Built-in training
3. Neither semantic nor API signals indicate training → General purpose

Discovery Variables

Variable	Tag	Notes
Experiment type	[depends]	Auto if signals are clear; ask if ambiguous
Parameters to sweep	[depends]	Auto-infer from hardcoded values; ask if unclear
Parameter values	[auto]	Suggest 2-3 alternatives around hardcoded values
Init function needed	[auto]	Yes if expensive ops independent of params detected
Function signature	[auto]	Determined by type (see REFERENCES.md)

Exit criteria: type classified, parameters identified, outputs known.

---

Phase 2: Design

Generate the experiment function and prepare the full design for user review.

1. Description — Create a 2-4 line description covering: what the experiment does, parameters being swept, outputs captured per trial, and evaluation criterion.

2. Parameters — Build the parameter table (Name | Values | Description).

   • Hardcoded value → suggest 2-3 alternatives around it
   • Categorical → list reasonable alternatives
   • [auto] Warn if total trials > 50 (suggest reducing values)

3. Function signature — Select based on type:

   Type	Signature
   General purpose	[out1, out2, ...] = func(params)
   Built-in (Form A, targets in datastore)	[trainingData, net, lossFcn, options] = func(params)
   Built-in (Form B, targets separate)	[trainingData, targets, net, lossFcn, options] = func(params)
   Custom training	output = func(params, monitor)

   Last 3 outputs for built-in must always be net, lossFcn, options. See references/REFERENCES.md for Form A vs B selection guide and supported loss functions.

4. Generate experiment function — Adapt user's code into the experiment function.

   Heading convention when presenting:

   • General purpose → "Generated Experiment Function"
   • Built-in training → "Generated Setup Function"
   • Custom training loop → "Generated Training Function"

   Key rules:

   • Use params.<field> for all tunable values (never hardcode swept values)
   • Each params.<field> is a single scalar value (numeric, string, or logical) — Experiment Manager assigns one value per trial. Never loop over parameter values inside the function; the sweep is handled externally.
   • Use 'Plots', 'none' in trainingOptions
   • Include H1 help comment block
   • Use absolute paths for all file/data references — see REFERENCES.md § "Absolute Path Rules"
   • For custom training monitor pattern — see REFERENCES.md § "Custom Training Monitor Pattern"

   Type	Body contains	Returns
   General purpose	User's computation logic, including figure/plot code	Named outputs (any serializable type)
   Built-in training	Data loading, network definition, trainingOptions	[data..., net, lossFcn, options]
   Custom training	Training loop with monitor integration, including figure/plot code	output (any serializable type)

5. [auto] Initialization function — If expensive one-time operations independent of hyperparameters are detected (data downloads, large file loads, datastores not depending on params), extract into an initialization function. See references/REFERENCES.md § "Initialization Function Template" for the code template.

---

Phase 3: Confirmation

Gate: [user] — Present design for review before execution.

MANDATORY: Always present the design and wait for explicit user confirmation before creating any files or running any MATLAB code. Never skip this gate, even if the user has confirmed similar experiments before. The only exception is if the user explicitly says to skip confirmation (e.g., "just create it without asking").

Presentation order matters:

1. Generated experiment function code
2. Generated initialization function code (if applicable)
3. Summary table
4. Ask for confirmation

Do NOT show the summary table before the function code.

Summary table:

Topic	Details
Experiment type	<type chosen and why>
Parameters/Hyperparameters	<comma-separated list>
Total trials	<number>
Function name	<functionName>
Initialization function	<initFunctionName> if generated, otherwise reason why not
Key metric	<metric name> (<minimize/maximize>)
Logged live metrics	<comma-separated list>

Proceed to Phase 4 only after user confirms.

---

Phase 4: Creation & Validation

Execute the experiment setup in MATLAB.

4a: Project Setup [auto]

1. Check currentProject — if open, use it; otherwise create a new project
2. Never close or replace an open project. If currentProject succeeds, always use that project — write the experiment function and .mat file into its root folder. Only call matlab.project.createProject when no project is open.
3. matlab.project.createProject changes pwd — use pwd directly after, don't prepend folder
4. Write function file into project root using writelines AFTER project exists
5. Generate unique folder names (append _2, _3) when creating new projects

See references/REFERENCES.md § "Step 6 Code Template" for the full code.

4b: Create Experiment Object [auto]

Use the createExperiment helper (in scripts/). It handles:

• Building the hyperparameter table from a params struct
• Generating unique .mat file names (never overwrites)
• Setting experiment name from file name
• Adding to project

4c: Validate [auto]

Before running validation, briefly tell the user what it checks: "I'm validating that the function exists on path, passes static analysis, has the correct signature, and that its parameter references match the hyperparameter table."

Path portability check: Before validation, scan for relative file paths in the generated function. If found, resolve to absolute and rewrite.

Run tiered validation (see REFERENCES.md § "Validation"):

1. Tier 0 — Existence: Confirm function is on path via which()
2. Tier 1 — Static analysis: Run checkcode(). Block on severity > 1 (errors). Warn on severity ≤ 1.
3. Tier 2a — Signature: Verify nargin and nargout match the experiment type:
   • general: nargin=1, nargout≥1
   • builtin_A: nargin=1, nargout=4
   • builtin_B: nargin=1, nargout=5
   • custom: nargin=2, nargout=1
4. Tier 2b — Parameter coverage: Parse function source for params.<field> references. Fail if function references params not in the hyperparameter table. Warn if table entries are unused.

Gate: [auto] — Validation result

• All tiers pass → open Experiment Manager (see REFERENCES.md § "Open Experiment Manager")
• Blocking failure → fix, rewrite .m and .mat, re-validate (max 2 attempts)
• Fail after 2 attempts → [user] escalate with error details

---

Phase 5: Handoff

After the experiment is created and Experiment Manager is open:

Your experiment is ready in Experiment Manager! Here are some things you can do before running:

Suggested authoring actions:

• Edit hyperparameter values or add new parameters in the Hyperparameter Table
• Modify the experiment description
• Open and adjust the function from the Experiment Manager app:
  • General purpose → "experiment function"
  • Built-in training → "setup function"
  • Custom training loop → "training function"
• Add or configure an initialization function from the app UI

When you're ready, click Run in the Experiment Manager toolbar to start your trials.

---

Key Rules

1. Always create new files — never overwrite existing ones (append numeric suffix).
2. Write function files inside MATLAB — use writelines directly into project folder.
3. Preserve user logic — adapt code into experiment function; don't rewrite their algorithm.
4. No training-progress plots — use 'Plots', 'none' in trainingOptions (built-in training only). For general purpose and custom training, preserve the user's figure/plot code.
5. Use params struct — all tunable values from params, not hardcoded.
6. Project-based — experiments must live in a MATLAB project.
7. Unique names — generate descriptive, unique names for functions and experiment files. Use suffix Project for project names and Experiment for experiment names (e.g., AntennaProject, AntennaExperiment).
8. Never expose internal APIs — don't show template class names to the user.
9. Use absolute paths — all file references must use fullfile() with absolute path components.

See references/REFERENCES.md for template classes, signature mappings, and common hyperparameters.

---

Error Recovery

Error	Fix
Validation fails	Fix function, rewrite .mat, re-validate (max 2 attempts)
Relative path found	Resolve to absolute against source script directory, rewrite
Project not open & can't create	Ask user for project location
>50 trials	Warn, suggest reducing values
Ambiguous experiment type	Ask user to clarify
File already exists	Append numeric suffix (_2, _3, ...)
Path cannot be resolved	Ask user for the full path

---

Copyright 2026 The MathWorks, Inc.