AGENT SKILL

spinning-up-deep-rl

Knowledge base from "Spinning Up in Deep RL" by Joshua Achiam (OpenAI, MIT-licensed). Use when applying Achiam's frameworks for RL fundamentals and MDPs, the model-free algorithm taxonomy, policy gradient derivations, the six reference algorithms (VPG, TRPO, PPO, DDPG, TD3, SAC), debugging silently-failing RL code, or running rigorous multi-seed RL experiments.

License 许可
MIT
最近更新
2026-08-30
Stars
★ 1
Forks
0

01 前言 Front Matter

name
spinning-up-deep-rl
description
Knowledge base from "Spinning Up in Deep RL" by Joshua Achiam (OpenAI, MIT-licensed). Use when applying Achiam's frameworks for RL fundamentals and MDPs, the model-free algorithm taxonomy, policy gradient derivations, the six reference algorithms (VPG, TRPO, PPO, DDPG, TD3, SAC), debugging silently-failing RL code, or running rigorous multi-seed RL experiments.

02 资源文件 23

03 SKILL.md 正文

Spinning Up in Deep RL

Author: Joshua Achiam (OpenAI) | Source: spinningup.readthedocs.io, MIT | Chapters: 20 | Generated: 2026-08-25

How to Use This Skill

  • No argument — load the core frameworks below
  • A topic — ask about advantage function, target networks, entropy regularization; I resolve it through the Topic Index and read that chapter file
  • chNN — I load that chapter's summary
  • "what chapters do you have?" — the full index
/cs:spinning-up-deep-rl                      # core frameworks + chapter index
/cs:spinning-up-deep-rl entropy regularization   # topic index -> ch19, read that chapter
/cs:spinning-up-deep-rl ch09                 # one chapter summary

When you ask about something not in Core Frameworks, I read the relevant chapter file before answering rather than guessing from the index.


Core Frameworks & Mental Models

The RL problem

pi* = argmax_pi J(pi), where J(pi) = E_{tau~pi}[R(tau)]. Every algorithm approximates this; where it substitutes a different objective (a Bellman residual, a surrogate), that substitution is the source of its failure modes. Four value functions — V^pi, Q^pi, V*, Q* — all obey Bellman self-consistency, and a*(s) = argmax_a Q*(s,a) is why Q-learning is a viable family at all. Advantage A^pi(s,a) = Q^pi(s,a) - V^pi(s) is the relative-quality signal policy gradients run on. (ch07)

The two branching questions

Place any algorithm by asking: does it have or learn a model, and what does it learn (policy, Q-function, value function, model). That generates the whole landscape. (ch08)

Policy optimization vs Q-learning — the central trade-off

  • Policy optimization is principled: you directly optimize the thing you want. Stable and reliable. On-policy, so it cannot reuse data, so it is sample-hungry.
  • Q-learning only indirectly optimizes performance, by training Q_theta to satisfy a self-consistency equation. Many failure modes, so less stable. But substantially more sample efficient when it works, because it reuses everything.
  • Satisfying the Bellman equations well carries no guarantee of good policy performance.
  • The two are not exclusive — DDPG and SAC live between them deliberately. (ch08)

The policy gradient template

grad J = E[ sum_t grad log pi_theta(a_t|s_t) * Phi_t ]. Five valid choices of Phi_t: full return, reward-to-go, reward-to-go minus a baseline, Q^pi, and A^pi. All share an expectation and differ in variance. Two rules get you from the first to the last:

  • Don't let the past distract you — drop rewards obtained before the action. Those terms had zero mean and nonzero variance: pure noise.
  • Baselines — by the EGLP lemma, any state-only b(s) can be added or subtracted freely. The standard choice is V^pi(s_t), learned by MSE regression onto reward-to-go. (ch09)

The policy-gradient loss is not a loss function

Its data distribution depends on the parameters, and it does not measure performance even in expectation. Only at the current parameters, with data from those parameters, does it have the negative gradient of performance. You can send it to negative infinity while performance craters, and it usually will. Only average return means anything. (ch09)

Broken RL code almost always fails silently

It runs fine; the agent just never learns. Usually something is computed with the wrong equation, on the wrong distribution, or piped to the wrong place. If it doesn't work, assume there's a bug before touching hyperparameters. Debug by measuring everything and reading the code critically. The archetype is one missing squeeze: a [N] vs [N,1] shape mismatch is broadcast-compatible, raises nothing, and silently turns the Bellman backup into an [N,N] matrix. (ch10, ch12)

Learn by doing

Write your own implementations, shortest correct version of each, simplest algorithms first. VPG, DQN, A2C, PPO, DDPG, roughly in that order; ~250-300 lines each. Single-threaded before parallel. Iterate fast in simple environments — under 5 minutes turnaround at the debug stage. Do not attempt Atari or Humanoid before the toy task works. Read papers for their ablations and supplementary material, but do not overfit to paper details (the original DDPG's architecture, init scheme and batch norm are not strictly necessary) or to existing implementations (their abstractions serve reuse, not your single use case). (ch10)

Rigor: four standards

  1. Fair comparisons — tune the baseline as hard as your method; never handicap it.
  2. Remove stochasticity as a confounder — at least 3 seeds, 10 or more to be thorough. Two seed groups can produce curves that look like different distributions.
  3. High-integrity experiments — launch fresh final runs and precommit to reporting them. Tuning produces hypotheses; final runs produce conclusions.
  4. Check each claim separately — ablate every design decision. (ch10)

The safe-step family (on-policy)

VPG takes an unconstrained gradient step, so a single bad step can collapse performance. TRPO constrains the step in KL-divergence between policies, not distance in parameter space, then backtracking-line-searches until the exact constraint holds. PPO drops the constraint and instead clips the objective so the policy gains nothing by moving far, which is first-order, far simpler, and empirically at least as good. (ch14, ch15, ch16)

The overestimation family (off-policy)

DDPG amortizes the intractable continuous max_a Q(s,a) into a learned policy: max_a Q(s,a) ~= Q(s, mu(s)). It needs a replay buffer (licensed because the Bellman equation is indifferent to how data was collected) and target networks (because the target otherwise depends on the parameters being trained). Its failure mode is Q-value overestimation, which the policy actively exploits. TD3 answers with clipped double-Q, delayed policy updates and target policy smoothing. SAC adds entropy regularization, making the explore-exploit trade-off an explicit coefficient alpha. (ch17, ch18, ch19)


Chapter Index

# Title Key Frameworks
ch01 Introduction The missing middle step, Code Design Philosophy
ch02 Installation Install-then-verify, MuJoCo optionality
ch03 Algorithms: What's Included and Why The two lineages, on/off-policy trade-off, code template
ch04 Running Experiments One flag per kwarg, ExperimentGrid, save-dir suffixes
ch05 Experiment Outputs Tools not files, watch-then-measure
ch06 Plotting Results Performance alias, prefix autocompletion, seed averaging
ch07 Part 1: Key Concepts in RL MDPs, four value functions, Bellman equations, advantage
ch08 Part 2: Kinds of RL Algorithms Taxonomy, model bias, policy-opt vs Q-learning
ch09 Part 3: Intro to Policy Optimization Log-derivative trick, EGLP lemma, reward-to-go, baselines
ch10 Spinning Up as a Deep RL Researcher Learn by doing, three idea frames, four rigor standards
ch11 Key Papers in Deep RL 13-section topic map
ch12 Exercises Problem Set 1 and 2, the silent DDPG bug
ch13 Benchmarks The parity disclosure, family-specific metrics
ch14 Vanilla Policy Gradient The six-step loop
ch15 Trust Region Policy Optimization KL trust region, line search, conjugate gradient
ch16 Proximal Policy Optimization PPO-Clip, KL early stopping
ch17 Deep Deterministic Policy Gradient MSBE, replay buffers, target networks, polyak
ch18 Twin Delayed DDPG Clipped double-Q, delayed updates, target smoothing
ch19 Soft Actor-Critic Entropy regularization, reparameterization, squashed Gaussian
ch20 Logger, MPI Tools and Run Utils EpochLogger pattern, MPI PyTorch order

Topic Index

  • Advantage function ch07, ch09, ch14
  • Baselines ch09
  • Bellman equations ch07, ch17
  • Benchmarks / parity ch13, ch01
  • Clipped double-Q ch18, ch19
  • Continuous action spaces ch07, ch17
  • Debugging / silent failure ch10, ch12
  • DDPG ch17, ch03, ch08
  • Entropy regularization ch19
  • Exploration vs exploitation ch14, ch17, ch19
  • GAE ch09, ch14
  • Installation ch02
  • KL divergence / trust region ch15, ch16
  • Logging ch20, ch05
  • MDPs ch07
  • Model-based RL ch08
  • MPI / parallelization ch20, ch02, ch04
  • MSBE ch17
  • Off-policy ch03, ch08, ch17
  • On-policy ch03, ch08, ch14
  • Papers / literature ch11, ch10
  • Plotting ch06, ch13
  • Policies (categorical, Gaussian, squashed) ch07, ch19
  • Policy gradient derivation ch09
  • PPO ch16, ch03
  • Q-learning ch08, ch07
  • Replay buffer ch17
  • Reparameterization trick ch19, ch10
  • Research process / rigor ch10, ch13
  • Reward-to-go ch09
  • Running experiments ch04, ch05
  • SAC ch19, ch03, ch08
  • Seeds / variance ch10, ch13, ch04
  • Target networks / polyak ch17, ch18
  • TD3 ch18, ch12
  • TRPO ch15, ch03
  • Value functions ch07, ch09
  • VPG ch14, ch09

Supporting Files

Scope & Limits

Covers the Spinning Up documentation only, as of the January 2020 PyTorch update. It does not cover: DQN and the discrete-action value-learning family (referenced, never implemented here), recurrent or convolutional architectures, partially-observed settings, model-based implementations, exploration/meta-RL/hierarchy beyond ch11's reading list, or any deep RL work after early 2020. The six implementations are educational; ch13 says which are research-grade. For topics beyond this source, I say so rather than improvising.


Compiled from OpenAI's Spinning Up in Deep RL documentation (MIT, Copyright (c) 2018 OpenAI), primarily developed by Joshua Achiam. Structured study notes, not a reproduction of the source.

04 同仓库更多 Skills

devlify-marketplace
devlify-marketplace/devlify-skills

syllabus

Generates a curated supplementary reading list from any course syllabus using Consensus academic search. Grill-me intake (syllabus input format + course audience + year range) plus a grouping forcing-options checkpoint before any search runs — so the reading list matches the course's level and recency need. Parses the syllabus to extract topics and learning outcomes, searches Consensus for recent peer-reviewed papers per topic, and produces a professionally formatted .docx with clickable Consensus links, plain-language summaries calibrated to audience level, and Bloom-higher-order discussion questions tied to course learning goals. Use when the user uploads a syllabus, course outline, or curriculum document and wants supplementary readings (e.g., 'create a reading list from this syllabus', 'find recent papers for my course') — even casual mentions with a syllabus attached should trigger this skill.

★ 1
devlify-marketplace
devlify-marketplace/devlify-skills

research

Default entry point for any research request — a hybrid router that classifies the question deterministically and either delegates to a specialist research skill (pulse for trends/sentiment, grants for NIH funding, litreview for academic literature, syllabus for course reading, patent for prior-art + IP landscape, dossier for entity research, deepread for evidence-first reading of supplied documents) or runs its own plan-decompose-multi-source-search-synthesize-cite fallback workflow when no specialist matches. Always surfaces the routing decision so users can override. Use when the user makes any research request that doesn't obviously match a more-specific specialist skill (e.g., "research [topic]", "look into [topic]", "what do we know about [topic]", "investigate [topic]", "find me information on [topic]", "do some research on [topic]", "I need to understand [topic]"). Output is a markdown briefing (default) or .docx document (on request) with full citations and an audit log.

★ 1
devlify-marketplace
devlify-marketplace/devlify-skills

pulse

Multi-source recency research skill that takes the pulse of any topic across Reddit, Hacker News, the open web, and optionally X/Twitter within a configurable recent window (default 30 days). Forcing intake clarifies topic specificity, angle (trend/sentiment/problems/opportunities/comparison), time window, and platform scope before searching. Returns a synthesized briefing with citations, engagement metrics, and cross-platform pattern analysis. Use when the user requests multi-source recency intelligence on a topic (e.g., 'pulse on [topic]', 'what's happening with [topic]', 'what are people saying about [topic]', 'current conversation about [topic]', 'take the pulse of [topic]', 'trending: [topic]', 'find me info on [topic]'), and for competitor research, trend discovery, tool comparisons, and audience sentiment analysis.

★ 1
devlify-marketplace
devlify-marketplace/devlify-skills

patent

Patent prior-art and landscape intelligence skill — not generic patent help. Commits to one of five sub-use-cases via forcing intake (novelty search / freedom-to-operate / competitive landscape / acquisition diligence / litigation prior-art) before any search runs. Searches Google Patents, Espacenet, USPTO, and optionally Lens.org for citation-graph signals. Output is an editable Word document (.docx) with verdict, ranked closest art (claim-text extracted), CPC-class-aware landscape, family-resolved hits, geographic coverage, FTO flags where applicable, strategy recommendations, and full audit log. Use when the user asks for patent searching or analysis (e.g., 'prior art search for [invention]', 'freedom to operate analysis for [product]'). Produces search signal, not legal advice — always recommends consulting a patent attorney before filing or licensing decisions. Trademark, copyright, and trade-secret questions are out of scope.

★ 1
devlify-marketplace
devlify-marketplace/devlify-skills

notebooklm

Browser automation skill for controlling Google's NotebookLM. Use when the user wants anything done in NotebookLM (e.g., 'open NotebookLM', 'check my [name] notebook', 'ask my notebook about X', 'add [source] to NotebookLM', 'generate a Video Overview from my notebook', 'use NotebookLM Studio'). Handles reading and querying notebooks, adding sources (URLs, text, files, YouTube links, synthesized content), generating Studio outputs (Audio/Video Overviews, Mind Maps, Reports incl. Briefing Doc/Study Guide/FAQ, Flashcards, Quiz, slide decks, infographics — discover the exact set from the live Studio panel; the UI evolves fast), and creating new notebooks. Requires browser automation environment — fails gracefully when unavailable.

★ 1
devlify-marketplace
devlify-marketplace/devlify-skills

litreview

Academic literature orientation skill that searches papers via free keyless APIs (PubMed E-utilities + OpenAlex) by default — with the Consensus MCP as an optional enhancement lane when connected — builds a strategic search plan using PICO (default) or SPIDER / Decomposition / hybrid as fallbacks, and synthesizes findings into a formatted Word (.docx) research guide. Grill-me intake (research question specificity + framework hint + tentative depth) before the recon search; a second forcing checkpoint after Phase 2 confirms framework + sub-areas + depth before searches consume budget. Configurable depth (5/10/20 queries) controls coverage vs. speed. Output is a 'launching pad' — an orientation guide that lets a researcher dive in confidently, not a finished review. Use when the user starts literature-oriented research (e.g., 'litreview on [topic]', 'literature review on [topic]', 'I'm starting a literature review on X', 'I'm writing a paper on X', 'help me research X', 'I'm doing research on X', 'can you help me research X'). Do NOT use for single one-off paper searches wanting a quick list — that's a plain PubMed/OpenAlex (or Consensus) query.

★ 1