2019-11-26 20:36:18
Sometimes with friends chat AI topic, I found that a lot of people on the AI situation has overvalued or undervalued excessive. Some questions do not need to use AI, some of the problems that the use of AI can not solve.
So I plan to write a series, from all angles to assess the question: "Do you want to use AI AI can solve my problem??"
Assessment angle: feature?
This article is cut angle: Features
Remember my junior high school teacher gave us talked about the evolution of a knowledge point: Russia perennial cold, Russian noses on the evolution of a very long, so that air entering the body needs to go a long way to the nose, not too cold a. So the Russians have a very obvious feature: the nose is long!
But not all big nose are Russians, Africans also have a big nose!
So, we wanted to determine whether the Russians need more features (evidence):
- Long nose
- tall
- Blue eyes
- White skin
- Orbital deeper
- Developed body hair
When we find that a person can have all the above features, then this person is the probability that the Russians would be much larger.
- This person can speak fluent Russian
When we found this feature (the evidence) above, which can be concluded that this person is Russian. Because of this feature is too strong, or too persuasive.
Spots - the basic principles of Artificial Intelligence
Recall the above process:
When we read a lot of Russians and people of other countries, we will summarize the experience of Russians characteristics: a long nose, high word, blue eyes, white skin, deep eye sockets, shaving developed, says Russian ...
When we encounter a foreigner not seen, we'll cover in this man with this "experience" to see whether, if a lot of features are met, it will guess this is Russian.
The basic principles of artificial intelligence is the above process, as shown below:
Characterized in Quadrant
But not all issues need to be resolved AI, AI's features can handle massive advantage, not only can handle the characteristics of the surface, hidden behind'll find features. But in many cases, no need to use artillery to fight mosquitoes.
When we draw a number of features and deterministic coordinate, guide what we will be able to issue suitable for AI, what the problem is not suitable for use:
+ Features less uncertainty weak: for artificial solve
+ Features a strong little certainty: the rules for solving
+ Features a strong multi-certainty: the rules for solving
Features more than + uncertainty weak: "consider" AI solve
PS: In fact, here there are many other factors: cost, risk, whether the measure ... here do not consider all, or too complicated.
Remarks
Upper quadrant too abstract, here cite a real case to illustrate.
The gaming industry there is a thing called "plug-in", simply plug is cheating, breaking the fairness of the game so that they have an advantage in the game.
Almost all well-known games have plug-in, plug-in because a lot of money! So game makers must be prepared to fight with the plug. Way against plug most of the cases are simple and crude, but effective: punished! Once the account using the plug-in, will handle title.
So the key is to strike the plug: the first time found using bots.
I have a friend in charge of a well-known hand combat tour, they also have a lot of plug-in, at the beginning they used some fixed rules to find the plug-in. The results were good, but there will still slip through the net.
So they try to use plug-AI task to do for a long time, it found that the effect is not much better than fixed rules.
Also against plug in some high complexity and high flexibility of the game (such as chicken, CS), not so that rule, because it is difficult summed up fixed rules.
This time AI will be able to show their talents. CS: GO and chicken are more successful cases of:
"Way he has applied Bishen - with plug-in machine learning ferret dog"
"AI in the game turned out the way these applications to see if you know a few? "
Quadrant application features:
Fighting games on the phone, freedom is not high, the player can only control movement, attack, skills, dodge, these four core operations. Strategic and not too strong, so roughly in line with "take the time to spend money + ≈ strength" logic.
You only need to grasp the principle: if the player did far more than their own strength to do?
By observing the players fighting, the enemy difficulty, time and other ways we can fight a more effective determine whether the player uses a plug-in. This can be a problem for people with a valid judgment, without the use of AI, but will complicate the solution.
Let us look at what wonderful plug-in chicken has: "chicken plug Guinness, one second allows you to see through each other directly Report! "
For CS: GO, chicken this shooting game, the scene is very complicated (large map, there is room, there is an obstruction ...)
Player's behavior is very complex (mobile, determine the position of the enemy, find cover, switch weapons, aim, fire ...)
In this case, it is difficult with clear rules to determine whether the player uses a plug-in. You can not say fast response even use plug-in perspective, some people react that fast; you can not say much on the use of headshot lock hanging, some people marksmanship is accurate.
So you want to find these plug-in, you need to analyze large amounts of data to find out the "not-so-obvious features" in the middle, this time AI had its special value.
to sum up
Today we are the "features" point of view which illustrates the problem suitable for AI, which is not suitable for the problem.
If the word is a word to summarize:
Can effectively summed up some rules questions, AI are not required, but is difficult to summarize those questions sum up the rules, consider using AI to solve.
If you want to evaluate the following features can be applied quadrant to see if your question is suitable for use AI techniques:
In addition to "feature" point of view, there are many ways you can help us determine: either use AI? This series will continue to update, follow me
