ClaytonStudent / MCTML_Light

Code for Master thesis "AutoML with Monte Carlo Tree Search and Neural Network"
2 stars 1 forks source link

Task 01.10-30.10 #5

Open ClaytonStudent opened 4 years ago

ClaytonStudent commented 4 years ago

01.10

Done

  1. 2个pipeline的构建,search space确定 ✅
  2. C值的动态变化,根据是否子节点全部被探索完毕来分配(是3,否10) ✅

To Do:

  1. Flexible 剪枝算法
  2. Fix & Flex 还可以增加的算法 分clf来训练 / NN 模型研究 / 取subset训练 ✅/ Node Selection时,最高值多于一个,选最近邻
ClaytonStudent commented 4 years ago

04.10

Done

  1. Flexible 剪枝算法 ✅ (需要修改,不根据sampling记录)
  2. MCTS 取subset (应用) ✅
  3. rollout取消掉✅
ClaytonStudent commented 4 years ago

05.10 Done

  1. 实现D-->F-->M-->H 的pipeline ✅
  2. 增加evaluator效率: sub-set ✅, history_sample✅
ClaytonStudent commented 4 years ago

06.10 Task

  1. 继续跑分,半小时起,保存模型和各种分数
  2. 修改MCTS,改进效率(查看论文,如何继续增加效率) 备选方案有 AMAF, RAVE, Progressive History, N-Gram Selection Technique. UCT,RAVE/AMAF,Progressive Bias,Virtual win & lose,Progressive Widening,LGR,Criticality
  3. IDEA: 不用神经网络预测分布,而是用每次跑完后Q值的比例 ✅
  4. 全的search space ✅ (需要调整冲突)
ClaytonStudent commented 4 years ago

07.10 Done

  1. 写开题报告,PPT ✅
ClaytonStudent commented 4 years ago

08.10 Task

  1. 把全的search space 冲突解决:如果节点是不能fit的,则删除该节点。✅
  2. 利用历史信息
ClaytonStudent commented 4 years ago

PPT: MCTS motivation / advantage / figure more / evalutation methods

Evaluations:

  1. statistical performance(柱状图 with or without one method,ranking / 点图),
  2. normalized difference,
  3. mean and standard deviationby repeated evaluation
  4. running time
ClaytonStudent commented 4 years ago

09.10 Done

  1. 改写 introduction ✅
  2. 阅读AlphaGo/ AlphaZero / AlphaD3M,总结NN的训练方法 ✅
  3. 搜集 Evaluation 方法 ✅
ClaytonStudent commented 4 years ago

10.10 Task

  1. 修改PPT 演讲稿
  2. 想出自己的NN训练法
  3. 写论文
  4. 跑数据
ClaytonStudent commented 4 years ago

15.10 TASK

  1. 修改PPT 演讲稿
  2. MCTS 跑起来 10min,30min 结果保存
  3. 目前MCTS有: sub-sampling/ history information / variable C / tree prune
  4. 目前MCTS缺少: variable probs(不用) /sample node expand(扩展新的节点) / inital state(暂时不用)
ClaytonStudent commented 4 years ago

论文可以加的内容:

  1. Progressive Widening (Expand new node) ✅
  2. Neural Network
  3. initial state classification ✅
  4. Flex pipeline
ClaytonStudent commented 4 years ago

21.10--24.10 Tasks

  1. 实现动态寻找初始的C值,如果可以随着时间变化,最后1min增大探索 ❌
  2. 动态调整P值(不必要) ❌
  3. 神经网络预测 performance(寻找之前的状态转化和代码)
  4. 剪枝算法,动态减去不需要的节点 ✅(同组最后一名,且与其他都差距巨大,visit time 足够大)
  5. 写论文,跑实验
  6. 每分钟保存最佳分数和配置 ✅
  7. 窗口 ✅ (目前random,需要推荐)
ClaytonStudent commented 4 years ago

25.10

  1. 不要纠结于C值, 简单介绍C值的取值范围 (不实现叙述)
  2. 神经网络重点在转化和架构 (实现,不介入)
  3. 窗口要有推荐值 (实现,接入)
  4. 树的剪枝算法 ( 不实现,叙述)
ClaytonStudent commented 4 years ago

26.10

  1. 基础table ✅
  2. 基础算法为代码 ✅
  3. 加入dataset 描述图 ✅
ClaytonStudent commented 4 years ago

28.10 Task

  1. 各个版本: 整理好作出表 ✅
  2. 实验计划: 表,期待的结果 ✅
  3. 论文 introduction

    一周之内 论文第一版

6-7 能跑的各个版本,上传到git和server 7-9 第一章开始写 9-11 第一章开始改

ClaytonStudent commented 4 years ago

28-31 Tasks

28: Introduction 29: Background 过一遍 30: Analysis 写完 31: Design 写完 32: Implementation / Conclusion 写完