Ptisni/cecm_precompute

Precomputed Experiment Artifacts

This dataset contains precomputed data used to reproduce cached experiments, planner behavior, and paper plots without rerunning expensive planning/training jobs.

Contents

precompute/
  exp_policies/
    level1-lot.npy
    ...
  rollouts/
    level1-lot.npy
    ...
  diverse_rollouts/
    level1-lot_0.35.npy
    ...
  trees/
    level1-lot.npy
    ...
  plans/
    level1-lot-1.npy
    ...

Files

exp_policies/*.npy

Precomputed Q-table policies for each level. These are used by the baseline agent and as action priors for MCTS planning.

rollouts/*.npy

Cached simulated trajectories for sampled capability vectors. These are used by CachedLearner when running standard cached capability inference.

diverse_rollouts/*.npy

Cached sets of diverse simulated trajectories for sampled capability vectors. Filenames include the diversity threshold, for example: level1-lot_0.35.npy. These are used by CachedLearner(diverse=True).

trees/*.npy

Cached state-transition trees for sampled capability vectors. These are mainly used for trajectory/tree validation and analysis.

plans/*.npy

Ground-truth planner rollouts used by experiment and plotting code. Filenames follow: -.npy For example: level1-lot-2.npy

How The Caches Are Used

Cached inference avoids rerunning MCTS simulations for every sampled capability vector. Instead, the learner loads simulated trajectories from rollouts/ or diverse_rollouts/ and compares them against observed trajectories. The planner policies in exp_policies/ are required by the baseline agent and by MCTS when using precomputed Q-value action biases. The plans/ files store reference trajectories and action plans generated under ground-truth capabilities. They are used by experiment scripts and plotting notebooks when comparing inferred/baseline behavior to ground truth.

Regenerating

These artifacts can be regenerated from the repository, but doing so may be expensive. ./get_rollouts.sh ./get_diverse_rollouts.sh ./get_trees.sh Q-table policies can be trained with: uv run python precompute/q_table.py --map-file level1-lot.txt --total-timesteps 100000 --q-default 0.01 Use the exact same environment version as the paper-release branch when regenerating artifacts.