agents-and-specifications
Use Tensorforce Agent.create, Agent.load, algorithm aliases, state/action specifications, action masking, and low-level act/observe or experience/update APIs.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- agents-and-specifications
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Use Tensorforce Agent.create, Agent.load, algorithm aliases, state/action specifications, action masking, and low-level act/observe or experience/update APIs.
- disable-model-invocation
- true
02 资源文件 6
- 参考文档references/agent-specifications.md9.2 KB
- 参考文档references/api-reference.md10.3 KB
- 参考文档references/interaction-apis.md7.6 KB
- 参考文档references/troubleshooting.md9.2 KB
- 脚本scripts/act_observe_smoke.py7.5 KB
- 脚本scripts/action_masking_smoke.py3.5 KB
03 SKILL.md 正文
Tensorforce Agents and Specifications
Use this sub-skill when a task asks how to construct or load a Tensorforce agent, choose an algorithm alias, write state/action specifications, pass action masks, or drive an agent without Runner through act, observe, experience, and update.
Route by task
- Create or load an agent: read
references/agent-specifications.md, then confirm exact public signatures and aliases inreferences/api-reference.md. - Choose between
ppo,dqn,random,tensorforce, and other aliases: use the alias table inreferences/api-reference.md. Route deep network, optimizer, memory, objective, policy, preprocessing, and parameter catalogs to../modules-and-configuration/. - Write state/action schemas: use
references/agent-specifications.md. Route custom environment implementations and adapter spaces to../environments-and-interaction/. - Use action masking: read the action-mask section in
references/agent-specifications.mdand run or adaptscripts/action_masking_smoke.py. - Write a manual training/evaluation loop: read
references/interaction-apis.mdand run or adaptscripts/act_observe_smoke.py. - Diagnose errors: use
references/troubleshooting.mdbefore changing dependencies or agent specs.
Key operating rules
- Prefer
Agent.create(..., environment=environment)when an environment object is already available; Tensorforce then infersstates,actions, andmax_episode_timestepswhere possible. - If no environment object is available, pass explicit
states,actions, and any required algorithm arguments. The primitive value spec isdict(type='bool'|'int'|'float', shape=..., num_values=..., min_value=..., max_value=...). - For non-independent training, every
agent.act(states=...)must be followed by exactly oneagent.observe(terminal=..., reward=...)for the same parallel interaction. - For evaluation, offline recording, or externally collected episodes, call
agent.act(..., independent=True)and useagent.initial_internals()when recurrent/internal states are present. - Action masks are auxiliary state entries, not part of
statesspecs. For a singleton integer action, useaction_mask; for a named integer action such asmove, usemove_mask. - Always close agents and environments you created directly. If
Runnerowns construction/closing, route to../runner-and-cli-workflows/.
Boundaries
This sub-skill owns agent creation/loading, algorithm alias selection, state/action specifications, action masking, and low-level interaction APIs. It intentionally excludes Runner orchestration and CLI workflows, custom environment implementation details, persistence/export/recording depth, and exhaustive module catalogs; use the sibling routes named above for those areas.
04 同仓库更多 Skills
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.
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.
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.
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.
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.
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).