Notebook Hyperparameter Pass (2026-04)¶
This pass tightens the default training budgets in the marimo notebooks for world-model training, opponent-model training, and RL experiments.
Main changes¶
- World-model notebooks now scale replay-buffer size and MLP capacity more explicitly with environment complexity instead of relying on one shared
train_world_model_mlp(...)default. - Gridworld world models use larger replay buffers and wider hidden layers because their public state is much larger:
pursuit:750ksteps for 2 agents,1.0Mfor 3 agents,hidden=1024gathering:750ksteps,hidden=768ice_duel:750kPPO-policy steps plus750kindependently seeded uniform-random steps,hidden=768markov_stag_hunt:750ksteps,hidden=512- Small matrix world models now use more data than the old
10kdefault, but still stay lightweight: bertrand,chicken,inspection:25ksteps,hidden=128,batch=256,epochs=80dpgg:100ksteps,hidden=256congestion:60kfor small agent counts,100kfor larger ones, with hidden size increasing to512for larger teams
Experiment-budget rationale¶
- Gridworld experiment notebooks now default to longer RL runs because they have larger observations and longer effective horizons:
pursuit:3.0Mgathering:3.0Mice_duel:2.5Mmarkov_stag_hunt:3.0M- Matrix experiment notebooks keep the same PPO-style update structure, but use longer runs and slightly smaller learning rates where the previous defaults looked too aggressive:
bertrand,chicken,inspection:750k,learning_rate=2e-4dpgg:1.0M,learning_rate=2.5e-4congestion:750kfor smaller agent counts,1.0Mfor larger ones,learning_rate=2.5e-4
Opponent-model changes¶
- Only the notebooks that looked clearly under-budgeted were changed.
dpggnow gives level-0 fitting more capacity and more imagined updates because the public pot state makes the repeated-game dynamics more stateful.inspectionnow uses a slightly wider level-0 network and more training updates because the role asymmetry makes small underfit errors more noticeable downstream.
What did not change¶
congestionstill uses imagination horizon1in opponent-model training. That is intentional: the observed state already summarizes the stage-level load outcome, so longer imagined horizons add cost much faster than they add signal.- The PPO rollout structure (
rollouts=2048,learning_epochs=4,mini_batches=4) was left alone for now. This pass focused on notebook defaults that were most obviously mismatched to environment size and horizon.
Follow-up if we want a more empirical pass¶
- Re-run the world-model notebooks and compare
wm_report.txtacross environments before increasing model width again. - Save level-0 holdout metrics from the OM notebooks to disk so future tuning can compare settings across runs instead of relying on notebook output.
- If matrix IPPO still shows seed instability after the longer budgets and lower learning rates, the next knob to try is a larger rollout size rather than another learning-rate drop.