[popular science] to understand supervised learning, unsupervised learning and reinforcement learning

1. introduction

Most introductory machine learning courses , At the beginning, the most basic concept is the three ways of machine learning , They are supervised learning (Supervised learning）、 Unsupervised learning (Unsupervised learning) And reinforcement learning (Reinforcement learning). Many students are unfamiliar with these three names . Here I try to explain the concept of parents by educating their children , If there is any omission , Welcome to give us more advice .

2. Machine learning is what the machine is learning

literally , machine learning Just let the machine （ The computer ） Learn one thing , Just like us , From small to large , We are deeply loved by our parents 、 Teachers, friends, etc , Formed the present us . Machine learning is the same , Let's assume that today's computer is your child , You can choose which educational policy to adopt to teach it , So that it can get the results you expect in the future .

3. Supervised learning

Let's first introduce our eldest son – Supervised learning . Its personality is simple 、 The steadfast , Most things we have to explain to it enough times first , It has enough judgment to make corresponding decisions .
for instance , We took it to the park today , To let him know what plants are , You point to banyan tree 、 Dwarf forest 、 Bushes and meadows , see ！ These are plants , Then point to the sky 、 Cars and houses say these are not plants , Take it to know most of the objects , The eldest son finally found the law （ green 、 Having roots or leaves ）, Learned how to judge what a plant is .

therefore , Supervised learning must be marked in the data （labeled） In the case of , In a real case , For large e-commerce , They may have a record of customers' monthly income 、 Age 、 Data list of gender and other markers , Feed the data to the computer for processing , Then the computer can judge when the next user appears according to the indicators , What is the probability that you will buy goods .

4. Unsupervised learning

Then let's introduce our second child – Unsupervised learning . It likes to classify the objects it sees according to their structures 、 Divide into different groups , When we put a box of animal toys in front of it , It will soon be able to separate different small circles , What is often surprising is , We didn't tell him in advance what category each animal belongs to , But it can be divided into winged ones by observation 、 An animal that can live in water or can only crawl on the road , Some even have their own unique classification methods that we did not expect , This is unsupervised learning . We don't have to go through supervision , The data can be classified from the toy by observing and analyzing the structure .

Unsupervised learning only needs to be unmarked （unlabeled） The data of , It can work normally . Replace it with a real example , What I think of is consumer preference analysis , Generally, when classifying different consumer groups , We are used to sex 、 Age and so on , But if we observe today that a member will buy cosmetics during the day 、 Buy beer in the evening 、 Buy an electric car in the evening , And the number of this population is not small , If you only look at gender and age, you should be confused , Do you think this consumer is suffering from schizophrenia . And this group of members is actually family , Family Dad 、 Mom 、 Children share the same account for online shopping , If we apply unsupervised learning to analyze consumer behavior , There is a chance to screen out these groups with the same consumption attributes .

5. Reinforcement learning

Finally our youngest daughter – Reinforcement learning . She likes playing chess best , We are often asked to play chess with it , At first we would snicker at its lack of common sense , Always make some obvious mistakes when playing dangerous chess , But after game after game , It constantly ponders whether the mistakes of each game can be changed , So the lower it goes, the better , The thinking routine is also deeper , you 're right ！AlphaGo It is an application of reinforcement learning .

A step closer , The characteristic of reinforcement learning is that training must have positive and negative returns ·（positive/negative reward）, In the process of training , The model will be based on different conditions （state） Try various decisions （action）, According to the result of this decision, we can learn and improve .

6. Conclusion

Finally, we make a simple summary of this article ：

• Supervised learning ： The data has been marked , Use the marked data for training .
• Unsupervised learning ： The data is not marked , Find data groups with the same characteristics .
• Reinforcement learning ： There may not be any data on hand , Let the model execute directly , Then feed back the execution results for training .

Boon boon , Have you learned to waste ？