AdapSafe2

This is the official implementation of AdapSafe2: Prior-Free Safe-Certified Reinforcement Learning for Multi-Area Frequency Control (https://ieeexplore.ieee.org/document/10726595).

This repository contains the core algorithm and multi-area frequency-control environments used for AdapSafe2.

Repository layout

• main.py: main training entry point.
• algs/MQL/: MQL / TD3-style meta-RL algorithm, buffers, parameter learning, and task snapshots.
• safe/: control barrier function (CBF) and safe critic utilities.
• maml/: MAML inverse dynamics model. Small default MAML checkpoints for two-area and N-area training are kept in maml/params_dir/.
• rlkit/envs/: two-area and N-area frequency-control environments plus normalization wrappers.
• models/: neural network modules.
• misc/: logging, runners, PyTorch utilities, and JSON/CSV helpers.
• configs/: environment/task configuration JSON files.

Setup

Python 3 is expected. The original codebase uses legacy Gym APIs, so a Python 3.7-3.9 environment is the safest starting point.

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Running

Example commands from the original experiment setup:

# AdapSafe2: safe critic + CBF + MAML inverse model
python main.py --enable_safe_critic --enable_cbf --enable_fake_transition --inverse_model MAML --env_name n_area_freq

# AdapSafe: CBF + DNN inverse model
python main.py --enable_cbf --inverse_model DNN --enable_fake_transition --env_name n_area_freq

# MQL baseline
python main.py --expl_noise 0.2 --env_name n_area_freq --lr 0.00005

By default, training writes checkpoints to ck/ and logs to log_dir/. These paths are intentionally ignored by git.