宁愿报虚警

1. Summary

1.1. Reinforcement Learning

1.1.1. Reinforcement Learning

1. Reinforcement learning tries all combinations cycling all data:
   • prove without any doubt
   • the best way is generally the objective that took the least amount of steps

QA may want to do 1 the first time as they’re learning as the code is discovering all the functionality of the application but after that they just want to use the best way usually, so after the system runs a neural network is populated with all of the information that neural network in essence is super user of QA’s application so QA can do some very impressive things with it —

• QA can tell it to run back all the tests using the most efficient way to solve each objective.
• QA could tell it to run all the positive every diffrent way that the objective succeeded in none of the negative ways.
• QA could tell it to run all the ways that failed
• QA can even pick a single thing and …

Requirements

• The objective must have measurable actions, that are readonably believed, in some combinaton, to be able to fulfill the objective.
• The actions must be executable automatically through a system that also has access to the results, and , can persist and update the persisted information.

1.1.2. 无监督学习

• 聚类
  • K-means: incresingly samller distances away surrounding points and the result is not controlled
• Neural Network¹: Every Little piece holds the data and its weight between 0~1 as a token of how confident it feels it is correct. What’s the most important,every singal point can benefit to the other points of data with their communication about the answer(?).--- The process must be helped with Intelligent Data Structure instead of traditional databases in face of massive queries & updates.

1.1.3. 有监督学习(Ensemble learning)

• 分类:知道了类标号,有输入输出
• 深度学习
• 决策树→随机森林

2. 实际应用

拿到标注好的良好的数据代价目前仍然挺高,比如医疗行业肿瘤图片数据一万五千笔近千万.面对这样的医疗数据,一定要把召回率也即检出率提上去:

• 数据查重,把阴性数据与阳性相仿的都改为阳性;宁愿报虚警

Footnotes

1. Using Selenium and AI to Create Test Cases ↩