Skip to content

README.md

Latest commit

 

History

History
45 lines (32 loc) · 1.9 KB

README.md

File metadata and controls

45 lines (32 loc) · 1.9 KB
Raw

Agentic-RAN Pipeline

Docker-first benchmark for slice-aware RAN forecasting and agentic control.

Corrected pipeline

docker compose up --build generate_data   # prepare dataset splits in shared_data/splits
docker compose up --build train           # train forecasting + agentic multi-task scenarios
docker compose up --build run-drl         # run offline DRL/agentic-policy control benchmark
docker compose up --build test            # aggregate final HTML report
docker compose up --build run-all         # generate_data -> train -> run-drl -> test

Fast development options:

DRL_EPISODES=5 DRL_MAX_SAMPLES=1000 docker compose up --build run-all
SKIP_DRL=1 docker compose up --build run-all

Outputs

  • Dataset splits: shared_data/splits/
  • Training artifacts: results/<scenario>/, results/train/train.html, ml_models/
  • DRL/agentic policy artifacts: results/policies/, results/tables/drl_seed_metrics.csv, results/tables/drl_slice_policy_kpis.csv
  • Final report: results/test/test.html and results/report.html

Project structure

The main implementation package is agentic_ran/. The folders models/, policies/, and src/ are kept only for backward compatibility with older tests/release scripts. See docs/PROJECT_STRUCTURE.md.

Agentic policy section

The report now contains a dedicated Agentic policy section. It includes:

  • whether agentic-policy features were enabled,
  • whether the model was an agentic action-head model,
  • action_accuracy,
  • action_macro_f1,
  • average decision confidence,
  • dominant recommended action,
  • DRL reward and slice-specific policy KPIs.

The action labels are pseudo-labels generated by an interpretable rule-based policy, not operator ground truth. They should be described as policy-consistency metrics. Control quality should be argued using DRL reward, slice KPIs, action-switch rate, safe fallback rate, and safety constraints.