Skip to content

ctde_v0

ctde_v0

Grounded CTDE v0 — the TeamBlue shared-exploration agent (multi-level L3/L1).

A from-scratch MAPPO-CTDE trainer on the zymera comm-coverage task, built to the TeamBlue architecture (agent_architecture.md / EXPERIMENT_PLAN.md), NOT a flat actor-critic:

  • config — the full §5-style nested run config (every knob).
  • nets — LPAC backbone (CNN -> GAP -> GNN message-passing KB with a configurable normalized aggregator) + goal-pointer / λ̂₂ / value heads (Actor) and the centralized Critic (CTDE).
  • controller — the L1 controller: goal stencil, greedy wall-aware move, and the action-mask mission-safety mechanism.
  • ppo — rollout (goal -> controller -> move), GAE on the central critic, PPO(goal) + aux-λ₂ + degree-reg + entropy, the trainer.
  • env_utils — env build, reward composition, true-λ₂ oracle, KB adjacency, degree stats, coverage metric.

Entry point: train_ctde.py.