Optienv

Optienv is a lightweight, simulator‑friendly framework for multi‑objective optimization (MOO). It wraps your external models with a simple CSV interface and lets you run evolutionary algorithms—currently NSGA‑II and NSGA‑III—at scale. It focuses on:

• Reproducible experiments: single‑file per‑seed histories, optional checkpoints/resume, and fixed seeds.
• Simulator integration: CSV adapter that reads variable assignments and writes objective values—no code rewrites inside your model.
• Many‑objective capability: NSGA‑III with reference directions for high‑dimensional objectives (≥ 4).
• Actionable analysis: global Pareto front extraction and normalized hypervolume (HV) computation without a user‑provided reference point, including density control via --epsilon.

---

Table of Contents

• Features
• Requirements
• Install
  • From PyPI
  • For Development
• Core Concepts
  • CSV Wrapper Contract
  • Run Configuration JSON
  • Single‑File History
• CLI Usage
  • search (NSGA‑II / NSGA‑III)
  • front
  • hypervolume (normalized, wide output)
• NSGA‑III Reference Directions
• Resume / Checkpointing
• Progress Output & Parallel Backends
• Examples
  • 3‑Objective Toy
  • 6‑Objective Toy (NSGA‑III vs NSGA‑II)
• Troubleshooting
• Roadmap
• License
• Citation / Acknowledgement

---

Features

• Algorithms
  • NSGA‑II: fast non‑dominated sorting + crowding distance (strong baseline).
  • NSGA‑III: fast non‑dominated sorting + reference‑direction niching (robust diversity for many objectives).
• CSV model adapter
  • Reads variable_values.csv (decision vector).
  • Calls your wrapper function (you control the simulation).
  • Reads objective_values.csv (objective values).
  • Converts maximize objectives to minimization internally (consistent sorting & indicators).
• Results management
  • One history CSV per seed (results/history_seed{SEED}.csv) containing all generations.
  • Optional final CSVs: population, fitness, and non‑dominated mask.
• Analysis
  • front: global non‑dominated set across all seeds + seed provenance column, optional ε‑box thinning.
  • hypervolume: normalized to [0,1] across all histories; no ref point needed; wide output (one HV column per seed).
• Robustness for Windows/Clusters
  • Lean worker processes (no heavy imports) and aggressive cleanup of temp folders (read‑only clearing, retries, background reaper, atexit).

---

Requirements

• Python: 3.10+
• OS: Linux / macOS / Windows
• Optional: pymoo (recommended for HV with > 2 objectives)

  pip install pymoo

---

Install

From PyPI

pip install optienv

Tip: Use a virtual environment (e.g., python -m venv .venv && source .venv/bin/activate on macOS/Linux or .venv\Scripts\activate on Windows).

For Development

git clone https://github.com/WRHGroup/optienv.git
cd optienv
pip install -e .

Reinstall after code changes to ensure the console script picks them up:

pip install -e .

---

Core Concepts

CSV Wrapper Contract

Your model’s wrapper (a Python file inside model_dir) must define:

def search_and_apply_variables(model_folder: str) -> None:
    ...

At runtime, Optienv:

1. Copies model_dir to a temp work folder.
2. Writes work/variable_values.csv (columns: Name,Value).
3. Imports your wrapper from the working copy and calls search_and_apply_variables(work).
4. Expects the wrapper to write work/objective_values.csv (columns: Name,Value).
5. Converts any maximize objectives to minimization (sign flip).
6. Deletes the working copy (unless --keep-work-on-error).

Tip (Windows/external tools)
Ensure your wrapper waits for external processes to finish and closes file handles:

import subprocess, os
with open(os.devnull, "wb") as devnull:
    subprocess.run(["your-sim.exe"], check=True, stdout=devnull, stderr=devnull, close_fds=True)

Run Configuration JSON

A minimal run_sim.json:

{
  "model": {
    "model_dir": "./toy_model",
    "wrapper_file": "wrapper_toy.py",
    "variables_csv": "../variable_declaration.csv",
    "objectives_csv": "../objective_declaration.csv"
  },
  "algorithm": {
    "population_size": 100,
    "generations": 200
  }
}

• model_dir: folder containing your wrapper file and any model assets.
• wrapper_file: Python file implementing search_and_apply_variables(...).
• variables_csv: CSV with columns Name,Upper_bound,Lower_bound.
• objectives_csv: CSV with columns Name,Objective where Objective∈{minimize,minimise,min,maximize,maximise,max}.

Single‑File History

results/history_seed{SEED}.csv (all generations in one file) with columns:

generation,index,<objective_1>,...,<objective_M>,<x1>,...,<xN>,nd

• Objectives appear first (either your names if --label-columns was used or fallback f0,f1,...).
• Variables follow.
• nd is a per‑generation non‑dominated mask (1 = non‑dominated among that generation’s population).

---

CLI Usage

search (NSGA‑II / NSGA‑III)

Run NSGA‑II:

optienv search \
  -c path/to/run_sim.json \
  --algo nsga2 \
  -j 4 --seed 7 \
  --label-columns \
  --no-save-final-csvs

Run NSGA‑III with reference directions:

optienv search \
  -c path/to/run_sim.json \
  --algo nsga3 --ref-parts 4 \
  -j 4 --seed 7 \
  --label-columns --no-save-final-csvs

Key options

• --algo: nsga2 (default) or nsga3.
• NSGA‑III only
  • --ref-parts <p>: build Das–Dennis simplex‑lattice reference directions (count = binom(M+p-1, p)).
  • --ref-dirs-csv <file>: supply custom reference directions (rows=dirs, cols=#objectives). Rows are auto‑normalized; overrides --ref-parts.
• Parallelism & infra
  • -j/--max-workers <N>: workers (processes by default).
  • --backend process|thread: use thread on RAM‑tight nodes.
  • --work-root <dir>: where temp model copies are created.
  • --keep-work-on-error: keep working copy if a failure occurs.
• Outputs
  • --label-columns: write objective/variable names as headers.
  • --no-save-final-csvs: only write the single history file.
• Resume/checkpoints
  • --checkpoint-every <N>: write a checkpoint every N generations.
  • --checkpoint-path <file>: where to store the checkpoint (default results/checkpoint.npz).
  • --resume-latest / --resume-from <file>: continue from a checkpoint.

---

front

Compute a global non‑dominated set across all results/history*.csv:

optienv front
# → results/pareto_front_all.csv

• Automatically converts objective sense using your declaration (if available).
• Adds a seed column (provenance).
• --epsilon <eps>: optional ε‑box thinning to control density (e.g., 0.01).

---

hypervolume (normalized, wide output)

Compute normalized HV per generation for each seed, and write a wide CSV (one HV column per seed):

optienv hypervolume
# → results/hypervolume.csv
# columns: generation, seed7, seed11, ...

• No reference point needed: objectives are first converted to minimization and normalized to [0,1] using global min/max across all histories; HV uses a fixed ref (r=\mathbf{1}).
• --epsilon <eps>: optionally thin each generation’s front first (speeds up HV, smoother curves). Try 0.01.

For objectives > 2, install pymoo to enable the HV indicator. Without it, a 2‑D fallback is used.

---

NSGA‑III Reference Directions

NSGA‑III uses reference‑direction niching (instead of crowding distance) to maintain a uniform spread across the Pareto front—especially important for many objectives.

• Default generation: Das–Dennis simplex‑lattice with --ref-parts p. Number of directions: [ \binom{M+p-1}{p} ]
• Best practice: population size ≈ number of reference directions (≈ one survivor per niche).
• High‑dimensional setups (e.g., (M\ge 6)):
  • Keep --ref-parts small (e.g., 3–4), or
  • Provide a custom --ref-dirs-csv with your desired count and distribution (e.g., 100–300 well‑spaced directions).

---

Resume / Checkpointing

Enable checkpoints during a run:

optienv search ... --checkpoint-every 1 --checkpoint-path results/checkpoint.npz

Resume later:

# Resume from default checkpoint
optienv search ... --resume-latest

# Or resume from a specific path
optienv search ... --resume-from path/to/checkpoint.npz

Checkpoints store generation index, population, fitness, RNG state, variable/objective names, bounds, history path, and model_dir—so resuming is safe and consistent.

---

Progress Output & Parallel Backends

During evaluation you’ll see:

[search] Gen 5/50 – evaluating 126 candidates (workers=8, backend=process)...
  completed:  63/126 ( 50%)  elapsed: 00:01:12
  completed: 126/126 (100%)  elapsed: 00:02:21
[search] Gen 5 done in 00:02:21

• --backend process (default) leverages multiple CPUs safely for external simulators.
• --backend thread can reduce memory overhead on RAM‑constrained Windows nodes.

---

Examples

3‑Objective Toy

Structure:

examples/toy_3obj/
├─ toy_model/
│  └─ wrapper_toy.py
├─ variable_declaration.csv
├─ objective_declaration.csv
└─ run_sim_example.json

Run:

cd examples/toy_3obj

# NSGA‑II
optienv search -c run_sim_example.json \
  --algo nsga2 -j 2 --seed 7 --label-columns --no-save-final-csvs

# NSGA‑III
optienv search -c run_sim_example.json \
  --algo nsga3 --ref-parts 8 -j 2 --seed 7 --label-columns --no-save-final-csvs

# Global front and HV
optienv front --epsilon 0.02
optienv hypervolume

6‑Objective Toy (NSGA‑III vs NSGA‑II)

Structure:

examples/toy_6obj/
├─ toy_model/
│  └─ wrapper_toy6.py
├─ variable_declaration.csv
├─ objective_declaration.csv
└─ run_sim_example.json

For M=6, --ref-parts 4 ⇒ 126 reference directions; match pop≈126 for NSGA‑III:

cd examples/toy_6obj

# NSGA‑II baseline
optienv search -c run_sim_example.json \
  --algo nsga2 -j 4 --seed 7 --label-columns --no-save-final-csvs

# NSGA‑III with matched pop size (edit JSON or override before running)
# Example using jq to set population to 126:
jq '.algorithm.population_size=126' run_sim_example.json > run_sim_example_126.json

optienv search -c run_sim_example_126.json \
  --algo nsga3 --ref-parts 4 -j 4 --seed 7 --label-columns --no-save-final-csvs

# Compare HV across seeds/generations
optienv hypervolume

You should see NSGA‑III achieve higher normalized HV and better spread in 6 objectives for a similar evaluation budget.

---

Troubleshooting

• Temp model copies persist on Windows

  • We clear read‑only flags and retry with exponential backoff; a background reaper thread and atexit attempt final cleanup.
  • Ensure your wrapper does not chdir, closes all file handles, and blocks for child processes to finish (subprocess.run(..., check=True, close_fds=True)).

• Many‑objective HV errors

  • Install pymoo for HV in ≥ 3 objectives:

    pip install pymoo

• Resuming with changed CSVs

  • The checkpoint validates variables, objectives, and bounds. If you changed them, start a fresh run.

• NSGA‑III population vs reference directions

  • It’s common to set population_size ≈ #reference_directions. If they differ you’ll get a note, but the run proceeds.

---

Roadmap

• More MOEAs: MOEA/D, IBEA, SMS‑EMOA (HV‑based survival).
• Reference‑direction factories (Riesz‑energy, multi‑layer designs) directly in the CLI.
• Rich plots & PDF reports.
• Surrogate‑assisted runs for expensive simulators.

---

License

MIT — see LICENSE.

---

Citation / Acknowledgement

If Optienv helps your research or engineering work, please cite or acknowledge it. This greatly helps the project grow and justifies continued development.

---