How qlt315/ucmec-mmwave-fronthaul works
Project type and purpose
• This is a multi-agent deep reinforcement learning (MADRL) evaluation framework for “User-Centric Mobile Edge Computing” (UCMEC).
• Its goal is to simulate decentralized task offloading and resource allocation in a user-centric MEC architecture, compare MADRL algorithms (MAPPO, IPPO, IQL, MADDPG) and convex optimization, and generate performance figures.
File structure
Top-level directories and files:
• algorithms/
– Contains MADRL algorithm implementations and helper modules.
– Subdirectories: algorithm/ (core policy classes: rMAPPOPolicy, r_actor_critic, r_mappo), utils/ (network modules: cnn, mlp, rnn; action/distribution classes; normalization).
• envs/
– Custom multi-agent environments for UCMEC under combinations of stat/dyna and coop/non-coop for different scenarios.
– Core env interface: env_core.py, with wrappers env_discrete.py and env_continuous.py.
• runner/
– Two runner styles: separated (each agent trained separately) and shared (agents share networks).
– base_runner.py and env_runner.py in each.
• train/
– train.py orchestrates experiment launch, parses --num_agents, chooses discrete vs continuous environment.
• utils/
– Shared utilities: replay buffers (separated_buffer, shared_buffer), normalizers, general util.
• scripts/
– data_process.py (post-processing), figure_generate.m (Matlab plots), render subdir with render_mpe.py and train_mpe.sh.
• config.py
– Hyperparameter and environment configuration.
• results/
– Output directory for logs, model checkpoints, metrics, and figures.
README for additional insights
• The README confirms the ENV abstraction (EnvCore, reset(), step()), continuous vs discrete action handling in algorithms/utils/act.py and runner/shared/env_runner.py.
• It points out how to switch scenarios by editing env files (N_sim, M_sim) and train.py flags.
Elements
– Config → train.py → Runner → Env (reset/step)
– Runner → Algorithm (action selection, learning updates)
– Algorithm → Utils (network forward/backward, buffers)
– Runner → Results (logging) → Scripts → Figures
Final advice
• Be exhaustive: represent every subcomponent (e.g., rMAPPOPolicy, shared_buffer).
• Branching logic: continuous vs discrete, shared vs separated training.