v0.5.3 — preserve the last finite Safe PINN training state

When a returned Safe PINN loss or gradient becomes nonfinite, or an Adam update produces nonfinite parameters or moments, the demo now saves the last completed finite training state and the failed batch in training-failure.mat, then rethrows the original error. Candidate Adam state is accepted only after these checks.

• Three controlled second-step regressions verify the exact preceding network, both finite moments, history and diagnostic inputs.
• All 107 local MATLAB tests passed, including 13 Safe PINN tests; the GitHub Actions run also passed. A separate two-update comparison with v0.5.2 preserved weights and numerical results, excluding timing.
• Only two of the 81 MATLAB files changed. The v0.5.2 records and prior release tags are preserved.

No new public API or automatic resume was added. Complete training, historical paper experiments, acceptance thresholds and scientific claims are unchanged.

Validation · Training-state check record

v0.5.2 — Retain rejected Safe PINN evaluations

The Safe PINN demo now saves its completed evaluation before checking held-out improvement and terminal accuracy. If a final check rejects the result, result.mat, metrics.json and training.csv remain available, and the console identifies their directory. The original assertion still throws.

• Added a real zero-update demo regression; all 104 local MATLAB tests passed.
• GitHub Actions passed repository checks and all 104 MATLAB tests on Ubuntu 24.04 with R2025b Update 6.
• A separate one-update success comparison preserved network parameters and numerical result fields, excluding timing. This uses its own curriculum and does not reproduce the 5,000-update experiment.
• The demo, its test and the test-runner description are the only changed MATLAB files. Network formulas, default settings, acceptance thresholds and historical experiments are unchanged.
• Original v0.5.1 validation records are archived, and the mean-field comparison remains tied to its tested source.

This patch preserves completed evaluations. Recovery from nonfinite losses during training or earlier evaluation errors is separate work. Full neural training was not rerun.

Validation and source identity · Inspecting a saved evaluation

v0.5.1 — Preserve mean-field failure data

The primary mean-field demo now saves its configuration, raw observations and statistics before policy fitting. Rejected fits retain replayable inputs and a small failure record while propagating the original learner exception. Successful runs append the learned result using the existing saved-variable names.

• Added one deterministic failure/replay regression; all 103 local MATLAB tests passed.
• GitHub Actions passed both repository checks and all 103 MATLAB tests on Ubuntu 24.04 with R2025b Update 6.
• A reduced successful run matched the preceding source for all saved numerical results, excluding elapsed time.
• Only the mean-field demo and its test file changed among the 81 MATLAB sources. Learner formulas, default settings and historical experiments are preserved.
• Includes the preceding baseline note and clearer commands for recorded experiments.

The v0.5.0 tag and its original test/source records remain available. This is a maintenance patch; the historical paper experiments were not rerun.

Validation and source identity · Failure replay instructions

ADP-MATLAB v0.5.0

ADP-MATLAB v0.5.0 is the first public release of a MATLAB reference project for adaptive dynamic programming control.

• 8 runnable entries and 81 MATLAB files: an integral-PI baseline and seven paper packages covering Koopman PI, mean-field LQG, off-policy Q-learning, HJB PINNs, safe control and Bias-PI.
• Bilingual documentation, equation maps, recorded numerical outcomes, editable figures and source data.
• 6 optional research skills for implementation, reproduction and extension.
• MIT License for independently written software and project documentation. External papers, upstream source and author checkpoints are excluded.

Open the repository root in MATLAB and run:

check_environment();
[result, runDir] = demo_reproductions('baseline');
results = run_all_tests;

The baseline uses base MATLAB with the standard JVM. The full test suite also requires Control System Toolbox and Deep Learning Toolbox.

Validation: all 102 local tests passed in two separate runs on MATLAB R2025b Update 6, macOS Apple silicon. The second run used a clean export of the 285 distribution files, without historical MAT data or author checkpoints. Both baseline quickstarts converged to the analytic LQR gain. Full neural training was not rerun.

The paper packages retain their documented numerical variants and limitations. Safe-PINN collision and budget violations, the robust-Koopman held-out error-bound failure, and Bias-PI performance gaps remain in the records. These code checks do not establish complete reproduction of the original papers.

Cloud CI: GitHub Actions passed both repository checks and all 102 MATLAB tests on Ubuntu 24.04 with MATLAB R2025b Update 6.

Project page · Getting started · Validation and scope