Model Parameter JSON

model_params_json lets Phase 6 pass model-specific settings into model wrappers. Use it when most STREAMLINE run settings can stay shared, but one model needs a specific constructor value.

The parameter is named model_params_json in .cfg files, notebooks, and the Phase 6 CLI.

How It Works

model_params_json is a mapping from model ID to parameter overrides:

{
  "MODEL_ID": {
    "parameter_name": "parameter_value"
  }
}

Model IDs are matched case-insensitively against the model wrapper small_name or full model_name. For example, HEROS, heros, and ExSTraCS are accepted model keys.

Use exact model wrapper parameter names. For constructor-controlled settings, such as HEROS and ExSTraCS hyperparameters, values must be present before the wrapper builds its Phase 6 parameter grid.

Expert Knowledge

Phase 6 automatically passes Phase 4 feature-importance scores into model wrappers that expose an expert-knowledge constructor parameter. ExSTraCS uses this through its expert_knowledge parameter.

The score vector is loaded separately for each CV split and ordered to match the features in that CV training file after feature selection. STREAMLINE uses Relief-style scores first when available (MultiSWRFDB, MultiSWRFDB*, MultiSURF, MultiSURF*) and falls back to mutual information.

You usually do not need to set expert_knowledge yourself. If you do supply it through model_params_json, your value takes precedence over the automatic Phase 4 score loading.

Config File Examples

In a .cfg file, place model_params_json in the [p6] section:

[p6]
outcome_type = Binary
models = HEROS,ExSTraCS
model_params_json = {"HEROS": {"iterations": 200000, "pop_size": 2000, "model_iterations": 1000, "model_pop_size": 200, "nu": 1}, "ExSTraCS": {"iterations": 200000, "N": 2000, "nu": 1, "rule_compaction": "QRF"}}

This runs HEROS and ExSTraCS with fixed values. It does not create a broad hyperparameter sweep because every supplied value is a single numeric or string value.

CLI Examples

For a direct Phase 6 CLI run, quote the parameter string:

python -m streamline.p6_modeling.p6_cli \
  --output_path out \
  --experiment_name DemoBinary \
  --outcome_type Binary \
  --n_splits 3 \
  --models HEROS,ExSTraCS \
  --model_params_json '{"HEROS": {"pop_size": 2000, "model_pop_size": 200}, "ExSTraCS": {"N": 5000, "nu": 1}}'

Use None when you want to restore a built-in candidate list for one parameter. This example searches only HEROS pop_size:

python -m streamline.p6_modeling.p6_cli \
  --output_path out \
  --experiment_name DemoBinary \
  --outcome_type Binary \
  --n_splits 3 \
  --models HEROS \
  --model_params_json '{"HEROS": {"pop_size": None}}'

You can combine fixed values and None in the same CLI value. This keeps HEROS mostly fixed while searching pop_size, and keeps ExSTraCS mostly fixed while searching N:

python -m streamline.p6_modeling.p6_cli \
  --output_path out \
  --experiment_name DemoBinary \
  --outcome_type Binary \
  --n_splits 3 \
  --models HEROS,ExSTraCS \
  --model_params_json '{"HEROS": {"iterations": 100000, "pop_size": None, "model_iterations": 500, "model_pop_size": 100, "nu": 1}, "ExSTraCS": {"iterations": 200000, "N": None, "nu": 1, "rule_compaction": "QRF"}}'

To run the built-in candidate sweep for every HEROS and ExSTraCS wrapper parameter, set each tunable parameter to None:

python -m streamline.p6_modeling.p6_cli \
  --output_path out \
  --experiment_name DemoBinary \
  --outcome_type Binary \
  --n_splits 3 \
  --models HEROS,ExSTraCS \
  --n_trials 200 \
  --timeout 900 \
  --model_params_json '{"HEROS": {"iterations": None, "pop_size": None, "model_iterations": None, "model_pop_size": None, "nu": None}, "ExSTraCS": {"iterations": None, "N": None, "nu": None, "rule_compaction": None}}'

Notebook Examples

In a notebook parameter block, use a Python dictionary:

P6_MODEL_PARAMS_JSON = {
    "HEROS": {
        "iterations": 200000,
        "pop_size": 2000,
        "model_iterations": 1000,
        "model_pop_size": 200,
        "nu": 1,
    },
    "ExSTraCS": {
        "iterations": 200000,
        "N": 2000,
        "nu": 1,
        "rule_compaction": "QRF",
    },
}

Use None in notebooks when you want the built-in candidate list for a specific parameter:

P6_MODEL_PARAMS_JSON = {
    "HEROS": {"pop_size": None},
    "ExSTraCS": {"iterations": None},
}

HEROS

HEROS uses these Phase 6 wrapper parameters:

Parameter	Fixed default	Candidate list when set to `None`
`iterations`	`100000`	`[100000, 200000, 500000]`
`pop_size`	`1000`	`[1000, 2000, 5000]`
`model_iterations`	`500`	`[500, 1000]`
`model_pop_size`	`100`	`[100, 200]`
`nu`	`1`	`[1, 10]`

Use pop_size for the top-level HEROS population size. Use model_pop_size for the internal model population size. HEROS does not use an N alias in STREAMLINE.

Example fixed HEROS settings:

{
  "HEROS": {
    "iterations": 200000,
    "pop_size": 2000,
    "model_iterations": 1000,
    "model_pop_size": 200,
    "nu": 1
  }
}

Example HEROS run where only pop_size is searched:

{
  "HEROS": {
    "iterations": 100000,
    "pop_size": None,
    "model_iterations": 500,
    "model_pop_size": 100,
    "nu": 1
  }
}

ExSTraCS

ExSTraCS uses these Phase 6 wrapper parameters:

Parameter	Fixed default	Candidate list when set to `None`
`iterations`	`200000`	`[100000, 200000, 500000]`
`N`	`2000`	`[1000, 2000, 5000]`
`nu`	`1`	`[1, 10]`
`rule_compaction`	`"QRF"`	`[None, "QRF"]`

ExSTraCS uses N for population size. HEROS uses pop_size.

When Phase 4 feature-importance results exist, STREAMLINE automatically passes CV-specific expert knowledge into ExSTraCS. Manual expert_knowledge values in model_params_json override the automatic Phase 4 scores.

Example fixed ExSTraCS settings:

{
  "ExSTraCS": {
    "iterations": 200000,
    "N": 2000,
    "nu": 1,
    "rule_compaction": "QRF"
  }
}

Example ExSTraCS run where only N is searched:

{
  "ExSTraCS": {
    "iterations": 200000,
    "N": None,
    "nu": 1,
    "rule_compaction": "QRF"
  }
}

Using `None`

By default, HEROS and ExSTraCS use fixed single-value settings. This protects runs from accidentally launching expensive RBML hyperparameter searches.

Passing None restores the built-in candidate list only for that parameter. If any model parameter has a candidate list with more than one value, Phase 6 uses Optuna for that model within the configured n_trials and timeout budget.

For .cfg files:

model_params_json = {"HEROS": {"pop_size": None}}

To sweep all built-in HEROS and ExSTraCS wrapper candidate lists from a .cfg file:

model_params_json = {"HEROS": {"iterations": None, "pop_size": None, "model_iterations": None, "model_pop_size": None, "nu": None}, "ExSTraCS": {"iterations": None, "N": None, "nu": None, "rule_compaction": None}}

python -m streamline.p6_modeling.p6_cli \
  --output_path out \
  --experiment_name DemoBinary \
  --outcome_type Binary \
  --n_splits 3 \
  --models HEROS \
  --model_params_json '{"HEROS": {"pop_size": None}}'

For notebook dictionaries:

P6_MODEL_PARAMS_JSON = {"HEROS": {"pop_size": None}}

To sweep all built-in HEROS and ExSTraCS wrapper candidate lists from a notebook:

P6_MODEL_PARAMS_JSON = {
    "HEROS": {
        "iterations": None,
        "pop_size": None,
        "model_iterations": None,
        "model_pop_size": None,
        "nu": None,
    },
    "ExSTraCS": {
        "iterations": None,
        "N": None,
        "nu": None,
        "rule_compaction": None,
    },
}

Common Mistakes

Use model_params_json, not model_json_param.

Use HEROS pop_size, not N.

Use ExSTraCS N, not pop_size.

Use Python values such as None, True, and False in CLI strings, cfg files, and notebook dictionaries.