AgentSkillsCN

Reinforcement Learning Assistant

全面的强化学习学习助手，适用于课程学习与自主研习。当您学习Q-Learning、策略梯度、Actor-Critic、MDP、贝尔曼方程，或探索各类强化学习概念时，可使用此技能。它能帮助您：(1) 通过类比深入浅出地讲解概念；(2) 分析并调试代码；(3) 在不直接给出答案的情况下提供作业辅导；(4) 搭建实验装置并进行数据分析；(5) 为自我评估设计测验题目；(6) 整理知识要点，制作复习资料；(7) 提供项目规划与实施建议；(8) 阅读并理解研究论文。

SKILL.md
--- frontmatter
name: Reinforcement Learning Assistant
description: Comprehensive RL learning assistant for coursework and self-study. Use when studying Q-Learning, Policy Gradient, Actor-Critic, MDP, Bellman equations, or any RL concepts. Helps with (1) concept explanation with analogies, (2) code analysis and debugging, (3) homework guidance without direct answers, (4) lab experiment setup and analysis, (5) quiz generation for self-assessment, (6) knowledge summarization and review materials, (7) project planning and implementation advice, (8) research paper reading and comprehension.

Reinforcement Learning Assistant

Course Directory Structure

The RL course materials are organized in aisd/courses/rl/:

code
rl/
├── labs/              # Lab instructions and original files (.docx)
├── notes/             # Study notes generated from resources (lab1.md, etc.)
├── resources/         # Source materials (bilingual articles, PDFs, images)
├── code/              # Implementation code and experiments
├── assignments/       # Assignment files
├── quizzes/           # Quiz materials
├── slides/            # Lecture slides (.pptx)
├── schedule/          # Course schedule (.docx)
└── scripts/           # Utility scripts (scrapers, generators)

Key workflow:

  1. •resources/ contains source materials (articles, textbooks)
  2. •notes/ contains study notes generated from resources
  3. •labs/ contains lab instructions (both .docx and bilingual .md)
  4. •code/ is for student implementations

When generating study materials, use resources as the data source, not labs.

Core Workflows

Understanding Concepts

Ask for explanations at your level (beginner/intermediate/advanced). Request analogies for intuition, then mathematical formulations when ready.

Request 5W1H Analysis Table for any core concept to get structured understanding:

  • •Why (为什么): Purpose and motivation
  • •What (是什么): Definition and components
  • •How (怎么做): Implementation and algorithm
  • •When (何时用): Use cases and scenarios
  • •Where (在哪用): Application domains
  • •Who (谁使用): Target users and researchers

Example request: "Create a 5W1H table for Q-Learning" or "Explain the core concepts of Lab 1 in a 5W1H table"

For detailed concept explanations: See references/concepts.md

Analyzing Code

Share code snippets and specify what you want to understand (flow, design choices, optimizations). Ask about common pitfalls.

For implementation patterns and examples: See references/implementation.md

Completing Homework

Describe the problem and what you've tried. Ask for hints and debugging strategies, not solutions. Verify your approach before implementing.

Running Experiments

  1. •Define experiment goal
  2. •Set up baseline
  3. •Change one variable at a time
  4. •Analyze results systematically

For experiment templates: See references/experiments.md

Testing Knowledge

Specify topics, question types (conceptual/mathematical/coding/applied), and difficulty level.

Reviewing Material

Request summaries, flashcards, concept maps, or algorithm comparison tables.

For quick reference: See references/quick-ref.md

Planning Projects

Follow phases: Planning → Implementation → Experimentation → Analysis

For project ideas and structure: See references/projects.md

Reading Papers

Use three-pass approach: (1) Abstract/figures/conclusion, (2) Problem/solution/experiments, (3) Derivations/details

For key papers: See references/papers.md

Common Pitfalls

  • •Incorrect reward shaping
  • •Insufficient exploration
  • •Wrong discount factor
  • •Not normalizing observations
  • •Forgetting terminal state handling

Best Practices

  • •Start with tabular methods before deep RL
  • •Visualize Q-values, policies, learning curves
  • •Test incrementally
  • •Use established libraries (Stable-Baselines3, RLlib)
  • •Compare against baselines