Using the code

Everything runs from the repo root. Each script prints the JAX devices it uses (CPU locally, GPU on a cluster) and writes outputs under results/.

The environments (problem setups)

name	actions	reward
relay_discrete	Discrete(K)	1 - |a - θ| (graded tracking)
robot_consensus	Box (continuous)	exp(-error) (formation quality)

The solvers (baselines)

All in src/coadapt/solvers/, each an independent learner per agent: reinforce, ppo, grpo, trpo, recurrent (PPO with a self-contained GRU belief for partial observability).

Entry points

# train one solver on one env, save a learning curve
PYTHONPATH=src python3 scripts/run.py --env relay_discrete --solver ppo \
    --N 8 --seeds 16 --horizon 60 --iters 300

# sweep network size N (static vs the "wall" condition)
PYTHONPATH=src python3 scripts/sweep_N.py --env relay_discrete --Ns 3 5 8 12 16

# 3-D collapse surface over (delay, drift) per N
PYTHONPATH=src python3 scripts/surface_delay_nu.py --env relay_discrete

# run EVERY solver and emit a Markdown comparison table
PYTHONPATH=src python3 scripts/compare.py --Ns 4 8 16 --iters 400 --seeds 32

Example wrappers

The examples/ folder has one shell script per task (01_train_single.sh, 02_sweep_N.sh, 03_surface_delay_nu.sh, 04_compare_solvers.sh). See examples/README.md.

Arbitrary seeds & horizon

--seeds, --horizon, --iters, --N are all call-time arguments. The rollout is scan over time and vmap over seeds, so the same code scales on a GPU by raising these.