-
Notifications
You must be signed in to change notification settings - Fork 0
Open
Labels
researchUse the website to experiment and conclude the resultUse the website to experiment and conclude the result
Description
Goal
Transition from the role of "Tool Builder" to "Researcher". Utilize the features developed in Issues 1-4 to conduct the first systematic investigation into AI behavior.
Research Question
How does the amount of computation (measured in maxVisits) affect KataGo's judgment (Win Rate), intuition (Policy), and global assessment (Ownership) of a position?
Tasks
1. Select Case Studies and Design Experiment
- Identify 3-5 interesting and complex board positions. Suggestions:
- A complex middle-game fight (high tactical uncertainty).
- An opening position requiring global judgment (high strategic complexity).
- A famous historical move (e.g., AlphaGo vs. Lee Sedol, Game 2, Move 37).
- Systematically analyze these positions using the Comparative View (Issue 3) at different
maxVisitslevels (e.g., 100, 1,000, 10,000 visits).
2. Analyze and Observe
Focus on these key areas during analysis:
- Intuition vs. Calculation: How much does the final MCTS result differ from the raw Policy Network? When is the difference largest?
- Decision Shifts: At what point (visit count) does the AI "change its mind" about the best move?
- Global Assessment Evolution: How does the Ownership Map evolve as visits increase? Does the AI become more or less confident in specific areas?
3. Document Findings (The First "Research Note")
- Create a new markdown file in the repository (e.g.,
docs/RESEARCH_LOG_01.md). - Write the first "Research Note," detailing the experimental procedure, observations, and preliminary conclusions.
- Crucial: Include screenshots from the visualization tool to illustrate the findings.
中文详细讲解 (Chinese Explanation)
1. 目标与价值:从"工程师"到"研究者"的关键跃迁
这是前 6 周计划的第一个高潮,也是你作为"AI 哲学家"的首次实践。在这个 Issue 中,重点不再是写代码,而是 提出问题、设计实验、分析结果并得出结论。
2. 行动步骤详解
-
设计实验: 利用你之前开发的工具。选择几个有挑战性的局面(案例)。使用"对比视图"(Issue 3),系统地改变
maxVisits(Issue 2),并观察"深度可视化"(Issue 1)的变化。 -
分析与洞察(最关键): 仔细观察数据,寻找规律和有趣的现象。你需要尝试回答"为什么"。
- 例如,你可能会发现:"在处理局部死活时,AI 的直觉(Policy Network)非常准确,增加计算量几乎不会改变结论;但在判断全局厚薄时,增加计算量会剧烈改变 AI 的世界观(Ownership Map)。"这就是一个有价值的洞见。
-
知识沉淀(撰写研究笔记):
- 极其重要。 将你的发现用清晰的语言和图表(直接从你的工具截图)记录下来。这是学术沟通的核心训练。
- 这篇文档的价值非常高,它将是你未来申请文书和潜在学术论文的直接素材。它证明了你不仅能创造工具,更能利用工具探索前沿问题并产生深刻见解。
Metadata
Metadata
Assignees
Labels
researchUse the website to experiment and conclude the resultUse the website to experiment and conclude the result