Sometimes I talk to friends about the topic of AI. I find that many people have over- or under-estimated AI. Some problems do not require AI at all, and some problems cannot be solved even with AI.
So I plan to write a series to evaluate a question from various angles: "Would you like to use AI? Can AI solve my problem?"
List of article series:
Evaluation perspective: characteristics?
The angle of this article is: Features
I still remember my junior high school teacher told us an evolutionary knowledge point: Russia is always cold, and the nose of Russians has evolved very long. In this way, the air entering the body needs to go a longer way in the nose, and it will not be too cold Already. So the Russians have a very obvious characteristic: they have long noses!
But not all noses are Russians, and Africans have noses!
So, if we want to judge whether we are Russian, we need more features (evidence):
- Long nose
- Blue eyes
- White skin
- Eye sockets are deeper
- Developed body hair
When we find that a person has all the above characteristics at the same time, then the probability that the person is Russian is much greater.
- This person can speak fluent Russian
When we find the above feature (evidence), we can basically conclude that this person is Russian. Because this feature is too strong, or too convincing.
Insertion-the basic principles of artificial intelligence
Review the above process:
When we have seen many Russians and people from other countries, we will summarize the characteristics of Russians based on experience: long nose, tall characters, blue eyes, white skin, deep eye sockets, well-shaved hair, speaking Russian...
When we meet a foreigner we haven't met, we use this "experience" to put it on this person to see if it fits. If many characteristics match, then we will guess that this is a Russian.
The principle of artificial intelligence is basically the above process, as shown below:
Feature quadrant map
But not all problems need AI to solve. The advantage of AI is that it can deal with a large number of features, not only the features on the surface, but also the hidden features behind it. But in many cases, it is not necessary to hit the mosquito with a cannon.
When we draw a coordinate of the number of features and certainty, we can guide us on what problems are suitable for AI and what problems are not suitable for use:
Less features + weak certainty: suitable for manual solution
Less features + strong certainty: suitable for rule solving
More features + strong certainty: suitable for rule solving
Many features + weak certainty: "can be considered" AI solution
PS: There are actually many other factors here: cost, risk, whether it is measurable... I don't consider it here, otherwise it is too complicated.
The upper quadrant is too abstract. Let's take a real case to illustrate it.
There is a kind of thing called "plug-in" in the game industry. Simply speaking, plug-in is a cheat device, which breaks the fairness of the game and gives yourself an advantage in the game.
Almost all well-known games have plugins, because plugins are very profitable! So game manufacturers must be prepared to fight with plug-ins. The method of cracking down on plug-ins is simple and rude in most cases, but it is very effective: severe punishment is not applied! Once the account is found to use a plug-in, it will be blocked for processing.
So the key to cracking down on plug-ins is to find players using plug-ins the first time.
A friend of mine is responsible for a well-known fighting mobile game. They also have a lot of plugins. At first they used some fixed rules to discover plugins. The effect is not bad, but there will still be fish that miss the net.
So they tried to use AI to catch plug-ins, and after doing it for a long time, they found that the effect was not much better than the fixed rules.
It is also a fight against plug-ins. In some games with high complexity and flexibility (such as eating chicken and CS), the rules are not easy to use, because it is difficult to summarize fixed rules.
At this time, AI can show its strengths. CS: Both GO and chicken have been successful cases:
Application characteristic quadrant:
Fighting games on mobile phones do not have high degrees of freedom. Players can only control movement, attack, skills, and dodge, which are the core operations of 4. The strategy is not too strong, so it generally conforms to the logic of "spend money + spend time ≈ strength".
So only one principle needs to be grasped: Do players do things far beyond their own strength?
By observing the player's fighting power, enemy difficulty, and fighting time, it can be more effective to determine whether the player has used the plug-in. For problems that can be judged effectively by people, there is no need to use AI, but the solution will be complicated.
First let's take a look at what wonderful plug-ins are available for eating chicken: "Eat chicken plug-in, let you spot the other party and report directly!"
For shooting games like CS:GO and Chicken Eating, the scene is very complicated (the map is large, there are rooms, there are obstructions...)
The behavior of the player is also very complicated (moving, judging the position of the enemy, finding cover, switching weapons, aiming, shooting...)
In this case, it is difficult to use clear rules to determine whether the player has used the plug-in. You can't say that the response is fast, even if you use the perspective plug-in, some people respond quickly; you can't say that there are too many headshots, and you use the lock hang, some people are accurate in shooting.
So if you want to discover these plug-ins, you need to analyze a lot of data and find "less obvious features" from the middle. At this time, AI has its special value.
Today we explained from the "feature" perspective which problems are suitable for AI and which are not.
In a word, it would be:
Problems that can effectively summarize some rules do not require AI, and those problems that are difficult to summarize rules can be considered using AI to solve.
If you want to evaluate, you can apply the following feature quadrant to see if your problem is suitable for AI technology:
In addition to the "feature" angle, there are many angles that can help us judge: Should we use AI? This series will continue to be updated, follow my public account to see everything:
Public number: Xiaoqiang-me