ctde_v0.train_ctde¶
Entry point: train the grounded CTDE v0 agent (LPAC + goal-pointer + λ̂₂).
Short CPU validation (16x16 / 4 agents): JAX_PLATFORMS=cpu /Users/bijanmehr/Project.Zymera/zymera_lab/.venv/bin/python -m ctde_v0.train_ctde --iters 20 --rollouts 8
Longer run (with a checkpoint + saved config sidecar): JAX_PLATFORMS=cpu /Users/bijanmehr/Project.Zymera/zymera_lab/.venv/bin/python -m ctde_v0.train_ctde --iters 200 --rollouts 16 --grid 16 --n-agents 4 --agg max --mechanism action_mask --run-dir runs/ctde_v0_16x16x4
Per iter it logs: episode reward, coverage %, connectivity %, mean λ₂, and the
aux-λ₂ accuracy (1 - median rel-err of λ̂₂ vs true λ₂), plus the controller
valid-move fraction (should be 100%). --run-dir writes config.json,
history.json and (if asked) a checkpoint next to the logs.