AGENT SKILL

environment-utils

Use PARL environment wrappers, replay buffers, schedulers, CSV logging, and summary helpers safely across Gym, continuous-control, Atari, MuJoCo, multi-agent, and vector-env workflows.

License 许可
Apache-2.0
最近更新
2026-09-03
Stars
★ 351
Forks
27

01 前言 Front Matter

name
environment-utils
license
Apache 2.0
metadata
{ "disco-role": "operating" }
description
Use PARL environment wrappers, replay buffers, schedulers, CSV logging, and summary helpers safely across Gym, continuous-control, Atari, MuJoCo, multi-agent, and vector-env workflows.
disable-model-invocation
true

02 资源文件 4

03 SKILL.md 正文

PARL environment utilities

Use this sub-skill when a task involves PARL's environment wrappers or small utilities around an RL loop: Gym API compatibility, continuous action remapping, Atari/MuJoCo/multi-agent/vector environment wrappers, ReplayMemory, schedulers, CSVLogger, logger, summary, tensorboard, or visualdl helpers.

Do not use this sub-skill to choose a full algorithm, implement a parl.Model/parl.Agent, or manage xparl workers. Route adjacent tasks to sibling operating skills:

  • ../core-framework/SKILL.md for backend selection, Model/Algorithm/Agent, save/restore, inference export, and weight synchronization.
  • ../algorithm-recipes/SKILL.md for DQN/DDPG/TD3/SAC/PPO/QMIX/MADDPG/A2C/IMPALA/CQL/OAC recipes and training-loop structure.
  • ../xparl-distributed/SKILL.md for @parl.remote_class, parl.connect, xparl start, remote environments, ports, and cluster security.

Operating workflow

  1. Normalize the environment API first. For ordinary Gym environments, wrap with CompatWrapper before calling code that expects the old four-return step API. Then add workflow-specific wrappers such as ActionMappingWrapper, wrap_deepmind, wrap_rms, MAenv, or VectorEnv.

  2. Choose wrappers by action/observation surface. Use ActionMappingWrapper only when the environment action space is a continuous Box; use Atari wrappers only for Atari-like image/action-meaning environments; use MuJoCo RMS wrappers for continuous-control normalization; use multi-agent wrappers only when the PettingZoo or legacy multiagent dependency stack is intentionally installed.

  3. Check data utility shapes before training. Configure ReplayMemory(max_size, obs_dim, act_dim) from the actual observation and action dimensions. For discrete actions pass act_dim=0; for continuous actions pass a positive integer action dimension.

  4. Keep logging output explicit. Set a deliberate logger.set_dir(...) before writing summaries, or create CSVLogger with an explicit file path in a scratch/output directory. Avoid accidental deletion from logger helpers when a directory already exists.

  5. Smoke-test utilities safely. Run the bundled checker before adapting examples:

    python scripts/check_env_utils.py --json
    

    Add --optional-wrappers to classify optional Atari, MuJoCo, and multi-agent imports without creating real environments.

  6. Escalate by symptom. For wrapper/data-flow behavior, read references/wrappers-and-data.md; for logging and schedulers, read references/logging-and-schedulers.md; for failures, read references/troubleshooting.md.

Verification status

The utility signatures for CompatWrapper, ActionMappingWrapper, VectorEnv, ReplayMemory, PiecewiseScheduler, LinearDecayScheduler, and CSVLogger were checked against an installed PARL 2.2.1 package. Small scheduler, replay-memory, CSV, vector-env, and action-mapping behaviors are covered by the bundled checker. Atari, MuJoCo, PettingZoo multi-agent, TensorBoard, and VisualDL flows are optional dependency surfaces and should be rechecked in the user's target environment before use.

Reference map

  • references/wrappers-and-data.md — Gym compatibility, action mapping, vector envs, Atari/MuJoCo/multi-agent wrappers, replay-memory contracts, and small NumPy RL helpers.
  • references/logging-and-schedulers.md — schedulers, CSV output, PARL logger directories, summary/TensorBoard/VisualDL behavior, and file-write safety.
  • references/troubleshooting.md — common fixes for Gym API mismatches, optional dependencies, action/replay shapes, logger deletion, CSV key mismatches, and summary backends.
  • scripts/check_env_utils.py — deterministic local checker for safe imports, signatures, and tiny pure-Python utility exercises.

04 同仓库更多 Skills

VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-triton-opt

Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-solve

END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-skill

PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-pattern-fusion

Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-orchestrate

Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-feedback

Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).

★ 351