Machine learning can be broadly divided into three categories: supervised learning, unsupervised learning, and reinforcement learning. Here, let's explain each mechanism.
① Supervised learning
Supervised learning is learning artificial intelligence in one direction by giving a set of examples and model answers (teacher signals). Generally, a large amount of data is required, and the neural network itself judges the correctness of the output result based on the given data.
It is possible to make judgments and actions by guessing from the examples even for cases that have not been learned, but it has the disadvantage that it cannot respond to unknown events that humans cannot be given knowledge in advance. In addition, there is a limit to the ability to "be smarter than the person who gave the model answer."
It is used for sales forecasting using "regression" that derives trends (functions) based on past data and predicts future numerical values, and image classification using "classification" that automatically classifies unknown data. increase.
② Unsupervised learning
Unsupervised learning does not require a model answer, and AI accumulates data based on its own activities and learns by itself.
It is a learning method that does not require a large amount of data, but instead, an "environment where you can learn correctly" is important. It is necessary to assume that the environment is consistent, and it is not possible to learn about events that cannot be simulated.
It is used for recommendations and customer segmentation utilizing "clustering" that analyzes accumulated data and extracts and groups similar ones from many.
③ Reinforcement learning
Reinforcement learning is a learning method in which AI repeats trial and error in its own environment to find the optimal behavior and value. In terms of recognizing and analyzing the results of AI's actions, it can be regarded as unsupervised learning.
An important factor in reinforcement learning is to make AI firmly aware of its own actions and situations. Then, the evaluation value for the result in the placed environment is used as a "reward" as a clue for learning.
For example, let's say you give AI an environment to play a game. Since there are no teachers, it does not show its strength at the beginning, but AI itself considers "how can I get more rewards" for each match. Data will be accumulated and become stronger with each battle.
In this way, reinforcement learning has a wide range of applications and is very effective when the object to be learned cannot be modeled.
