v1.5.0

Contextual Optimization (LCE-M GP)

• Added contextual multi-task optimization built on BoTorch's LCEMGP (Feng et al., NeurIPS 2020) for both the single-objective (bo.py) and multi-objective (mobo.py) BoTorch backends.
• Context embeddings are definable: learned from data, supplied manually per context (any encoder, e.g. ViT-G/14 vectors), or computed from context images via open_clip (default ViT-bigG-14, the open_clip release of ViT-G/14) with content-hashed caching and optional L2 normalization.
• Warm-start parameter CSVs accept a Context column to transfer observations from other contexts (users, devices, environments); new observations are tagged with the current context.
• Run metrics (coverage, IsBest/IsPareto, hypervolume/best-objective traces) and the logged Iteration index are computed over current-context observations only; ObservationsPerEvaluation.csv gains a Context column.
• New BoForUnityManager inspector section with live validation and fail-fast startup checks; contextual mode is BoTorch-only (CABOP rejects it with a clear error).
• Worked around a BoTorch LCEMGP task-kernel dimensionality issue when context_emb_feature is provided.

CABOP Fixes

• Fixed a critical parameter-ordering bug: with multiple CABOP groups whose parameters interleave in declaration order, parameter values were silently assigned to the wrong parameter names (in Unity payloads, observation logs, and warm-start data). Vectors now always follow parameter declaration order.
• Fixed a spurious AssertionError when the acquisition optimizer landed exactly on a parameter bound (floating-point overshoot); proposals are now clamped to bounds.
• IsBest/IsPareto marker flags are now derived from full-precision scalarized values instead of the rounded CSV values, and no longer re-scan the whole CSV every iteration.
• Zero-configured costs and degenerate GP predictions no longer produce division-by-zero in the acquisition function.
• The CABOP runtime now tolerates malformed/unrelated protocol messages the same way as the BoTorch backends.

Runtime and Tooling

• The optimizer's Python process now runs with PYTHONDONTWRITEBYTECODE=1, keeping __pycache__ folders out of StreamingAssets.
• Fixed a pandas dtype issue when writing IsBest flags during bo.py sampling runs.
• Context image paths configured on Windows now also resolve on macOS/Linux.
• MainThreadDispatcher no longer holds its queue lock while running actions, and one failing action can no longer abort the rest of the frame's queue.
• Removed the dead Optimizer.UpdateParameter API (it referenced CSV data that was never loaded).
• Added example warm-start CSVs with a Context column (ExampleContextInitData*.csv); documented the optional open_clip_torch/pillow dependencies in requirements.txt.
• Added a manually triggered Release workflow that tests, rolls this changelog, bumps bundleVersion, and creates the GitHub release.

Tests

• Added unit tests for the context protocol and embedding pipeline, plus real-BoTorch integration tests running full contextual BO/MOBO loops (skipped automatically on CI environments without torch).
• Added CABOP backend tests, including a multi-group ordering regression test and bounds/zero-cost edge cases (skipped when scipy/scikit-learn/loguru are unavailable).
• Added Unity EditMode tests for final-design selection and objective key matching.
• Consolidated the duplicated torch/botorch test stubs into a shared tests/_stubs.py.
• Added a weekly/manual full-stack-tests CI job that runs the complete suite against the real pinned dependency stack (CPU torch).

Documentation

• New README section 8.13 on contextual optimization and context embeddings (incl. ViT-G/14 guidance), new troubleshooting entries, and a note on Iteration numbering semantics for warm-started and contextual runs.

v1.4.2

Fitts Law Study Workflow

• Added a unified Fitts Law scene with selectable HITL MOBO, Static, and Random conditions.
• Added explicit Fitts Law design parameters: x_font_size, button_size, button_distance, button_hue, and button_saturation; brightness remains fixed at 0.5.
• Added target X marker support and disabled target outlines/red wrong-click flashing by default.
• Added layout safeguards to prevent target overlap.

Objectives and Logging

• Updated Fitts Law objectives to aesthetics, speed, accuracy, and usability.
• Logs now use a shared structure under Assets/StreamingAssets/BOData/LogData/<USER>/<CONDITION>/.
• Questionnaire logs now include UserID, ConditionID, and GroupID via Additional CSV Items.
• Added detailed Fitts app/trial telemetry logs, including click counts, timing, accuracy, iteration, phase, and design parameters.
• Added duplicate user-folder protection using suffixes like _1.

Runtime and Tooling

• Improved Python setup to avoid main-thread Unity API calls from worker threads.
• Updated Python requirement handling for the bundled Python 3.13 runtime.
• Fixed PlayMode test asmdef references and removed the previous UnityTest dependency issue.

Documentation

• Expanded README coverage for Fitts Law conditions, parameters, objectives, questionnaire setup, logging paths, and troubleshooting.

v1.4.1

• Replaced BoTorch’s internal Pareto/hypervolume metric utilities with moocore-backed wrappers while preserving existing MOBO behavior.
• Added tensor/array conversion helpers, maximization-aware Pareto and hypervolume wrappers, duplicate-first Pareto retention, updated tests, and moocore installation support for requirements plus Linux/macOS setup checks.
• Also improved Unity Python dependency setup by trying an upgrade install first, then falling back to a non-upgrade install so existing compatible packages could still be reused if upgrades failed.

v1.4.0

Added

• New Fitts' Law task for experimental and interaction evaluation workflows.

Fixes

• Various bug fixes with the Unity Editor inspector user interface.

Cleaned Up

• Updated the repository's .gitignore to better match Unity project structure.
• Removed generated IDE, solution, and Unity build/runtime artifacts from version control.

v1.3.0

🚀 Robust HITL BO Loop & Protocol Hardening

• Replaced ad-hoc socket reader with buffered NDJSON framing (recv_json_message), preserved unread bytes, added socket timeouts, SO_REUSEADDR, and stricter send/receive error handling.
• Added extensive fail-fast validation (parameters, objectives, bounds, shapes, non-finite values, duplicate keys, CSV layout checks).
• Enforced batchSize=1 (warning) and improved logging of init config, timeouts, and warm-start format.

📊 Warm-Start Reliability & Normalization

• Added WARM_START_OBJECTIVE_FORMAT (auto | raw | normalized_max | normalized_native) with strict validation.
• Improved normalize_param_column / normalize_obj_column:
• Auto-detect raw vs. normalized data with epsilon tolerance.
• Correct scaling to [0,1] (params) and [-1,1] (objectives).
• Proper sign handling for minimization flags.
• Safe clipping and degenerate-range handling.
• Hardened warm-start CSV loading: existence checks, column validation, numeric conversion, row consistency, model-space normalization.
• Preserve full precision when denormalizing optimizer parameters.

🔍 Data Integrity & Logging Improvements

• Runtime sanity checks for iteration counts, restart/sample sizes, objective validity.
• Improved objective reception: explicit errors for missing/non-finite values.
• Best objective now logged to BestObjectivePerEvaluation.csv (mirrored to legacy HypervolumePerEvaluation.csv).
• Robust IsPareto / IsBest handling with fallback logic and mismatch warnings.
• Improved CSV IO helpers and configuration parsing.

🔄 Iteration Flow Control (Unity)

• New iteration progression modes:
• NextButton
• ExternalSignal
• Automatic (with delay coroutine)
• Added RequestNextIteration() API.
• Improved state management, scene reload toggle, and protection against double-advances.
• Updated Inspector UI to expose new settings.

🏁 Optional Final Design Evaluation

• Added FinalDesignSelector:
• Deterministic selection from ObservationsPerEvaluation.csv
• Supports utopia distance, maximin, aggression metrics with tie-breaking.
• Integrated optional final round (totalIterations + 1) without sending objectives back to Python.
• Inspector controls and README documentation added.
• Input validation and error reporting for selection failures.

🧪 Tests & Documentation

• Added unit tests (test_bo.py, test_mobo.py).
• Expanded README: glossary, quick start, integration checklist, warm-start examples, troubleshooting.
• Improved diagnostics for protocol and normalization debugging.

🛠 Miscellaneous

• Added socket buffer constant and timeout env config (BO_SOCKET_TIMEOUT_SEC).
• Improved precision handling, tensor stacking, bounds validation.
• Updated Unity project files and platform scripts.
• Included StreamingAssets BO data and refreshed project metadata.

v1.2.0

Highlights:

• Added support for single-objective optimization.
• Added single-objective example scene: Color Guesser (3 Parameters, 1 Objective)

Full Changelog: v.1.1.2...v1.2.0

v1.1.2

What's Changed

• Guard success message on expected optimization completion

• Handle successful socket shutdown without noisy logs

• Clarify peer shutdown handling

• Upgraded Python version and Python library requirements

• Clarified how to get parameter values and set objective values via code on the Unity side in the README.

Full Changelog: v1.1.1...v.1.1.2

v1.1.1

What's Changed

• Harden questionnaire and network parsing against locale differences.
• Changed socket communication to JSON messages.

Full Changelog: v1.1.0...v1.1.1

v1.1.0

Highlights

• Click Game example scene
  Minimal 2-objective task (time-to-click, perceived difficulty). Ready to plug into the BO loop.
• Acquisition upgrade: qNEHVI → qLogNEHVI
  Replaced legacy qNoisyExpectedHypervolumeImprovement with qLogNoisyExpectedHypervolumeImprovement for better numerical stability. Same API.
  from botorch.acquisition.multi_objective.logei import qLogNoisyExpectedHypervolumeImprovement
• Simplified BO Manager inspector (Unity)
  Cleaner grouping, clearer hyperparameters, and safer defaults.
• Python bootstrap
  Auto-detects a Python executable and installs requirements.txt from
  StreamingAssets/BOData/Installation/ automatically.
• Sampling rule: 2d + 1
  The number of sampling iterations now defaults to 2 * num_design_params + 1. Toggle manual override if needed.

Upgrade notes

1. Ensure StreamingAssets/BOData/Installation/requirements.txt is present; the runner installs packages automatically.
2. If you previously set a fixed sampling count, enable “Set Sampling Iterations Manually” in the inspector to keep. Otherwise, the default is 2d+1.

Changelog

• Add: Click Game example scene.
• Change: Switch to qLogNoisyExpectedHypervolumeImprovement.
• Improve: BO Manager inspector UX.
• Add: Default Python path detection + automatic requirements.txt install.
• Add: Automatic sampling iterations = 2d+1.

v1.0.1

Now using the qNEHVI acquisition function as a default setting.

qNEHVI is far more efficient for noisy data typical of human feedback than qEHVI and mathematically equivalent in the noiseless setting, see: https://botorch.org/docs/tutorials/multi_objective_bo/