# mycode

**Repository Path**: a-baii/mycode

## Basic Information

- **Project Name**: mycode
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-06-12
- **Last Updated**: 2026-07-20

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

## EdgeFed H-MAAC: Edge Federated Heterogeneous Multi-agent Actor-Critic

This repository contains a Gymnasium-style UAV-assisted MEC simulation environment and a PyTorch migration of the historical TensorFlow EdgeFed H-MAAC implementation.

### Run

- To simulate the MEC system, the MEC world is implemented by `MEC_env/mec_def.py` and wrapped as a Gymnasium-compatible environment in `MEC_env/mec_env.py`.
- The active MAAC and AC implementations are PyTorch ports under `torch_agents/`.
- Historical TensorFlow reference code is archived under `legacy/tensorflow/`.
- Experiment entry points live under `runs/`. For example, run migration validation with `python -m runs.run_maac_reproduction`.
- Shared Python parameters live under `configs/Params.py`.

### Migration Status

- The current project goal is equivalent migration, not algorithm optimization.
- Active training code should preserve the TensorFlow training semantics unless a change is required for PyTorch or Gymnasium compatibility.
- `MEC_MARL_ENV.step(...)` returns `(obs, reward, terminated, truncated, info)`.
- `MEC_MARL_ENV.reset(...)` returns `(obs, info)`.
- Use `python -m runs.run_maac_reproduction` to run the full migration validation against the historical TensorFlow baseline. Pass `--tensorboard` only when an interactive TensorBoard view is needed.

### Validation

The historical TensorFlow baseline is `20260402-124320`. A PyTorch migration run is considered close enough when:

- last-100 average `step_reward` is no more than `125%` of the TensorFlow baseline.
- final `finish_len` is at least `95%` of the TensorFlow baseline.
- final `finish_data` is at least `95%` of the TensorFlow baseline.

`step_reward` is an AoI cost metric, so lower is better.

### Current Research Layout

The root is kept thin: project configuration, package directories, and
documentation only. Runnable training and validation entry points live under
`runs/`, and experiment or analysis scripts live under `scripts/`.

Experiment entry points are collected under `runs/`:

- `runs/main_run.py`: choose which algorithms to run for the current comparison round.
- `runs/run_maac_reproduction.py`: PyTorch migration validation against the historical baseline.
- `runs/AC_run.py`, `runs/MAAC_run.py`, `runs/H_MAAC_run.py`: original migrated algorithm entry points.

Current stable and baseline agents stay in `torch_agents/`. Older or exploratory
agent variants are kept importable under `torch_agents/experimental/`. Auxiliary
one-off tools live under `scripts/tools/`, while obsolete runnable scripts are
kept under `scripts/archive/`.

Experiment logs are organized under `logs/` with English directory names:

- `logs/main_algorithm/`: current main algorithm runs, useful archives, and deprecated attempts.
- `logs/energy_rework/`: energy-accounting and Pareto-related runs.
- `logs/graph_neural_experiments/`: graph-module experiments with active, promising, and deprecated buckets.
- `logs/mobility_rework/`: moving information-source and road-aware mobility runs.

Generated image assets live under `figures/`, including `figures/imagegen/`.

### References

* *Federated Multi-Agent Actor-Critic Learning for Age Sensitive Mobile Edge Computing*
* *An Edge Federated MARL Approach for Timeliness Maintenance in MEC Collaboration*