Can the AI learn from its mistakes?

That is a good question. The answer is simple. That depends on how the error or mistake is determined. Determining what the mistake or error is is one of the bottlenecks to creating artificial general intelligence (AGI): how to determine right or wrong? How to describe favorable cases. That the AI should use. And. How to determine the non-favorable cases? The latter are cases that the AI should avoid. In the teaching process. The operator determines and describes the case. And then gives it. Positive (favorable) or negative (non-favorable) values. 

In this text. The main topic is reinforcement learning. There, the system makes something, and the environment. It gives feedback. The feedback. Or. The actor who gives feedback. Determines. If the AI acts right or wrong. And how to determine the values “yes” (Positive)(+)  and “no” (Negative)(-)? 

The system could partially follow Boolean algebra. The chain of positive (+) solutions. It can be conjucted by using “AND”. The “OR” changes the model. And if the model gives a negative (-) value. The system turns to using “NOT,” and then the system. It must change the model. The disjunction happens when there are too many negative values in the series of cases. The negation operation makes the system retake the algorithm. And then try another way, or algorithm, to solve the problem. The problem. It must always be solved by following the rules. 

When we think about the trial-and-error model. That model is effective. But not in all cases. This model is also known as the reinforcement model. Trial-and-error model. It is a good tool for virtual cases. But in cases where the AI must drive a car. That kind of learning solution. That can turn very expensive. There are not many ways. How to react to things the right way. Wrong reaction. It can turn fatal. If the AI driver reacts the wrong way. That can be a very big risk.  

When the car stops at a red light. That is the rule. This instruction is for public safety. But what if somebody tries to rob the car? What if a street gang member  shows a red light or “stop sign” to the car? Trying to rob it? That case is not very common. But those special cases show. How difficult. It is to program the AI. The AI is like a student. That system requires intensive training. The AI trainer must give instructions on what to do. And what not to do. 

In simple cases, the AI uses a limited data type. The AI is very easy to teach. The system requires a description of the favorable case. That case is determined as plus. But then the AI requires determination. About the non-favorable cases. The thing that the AI should not do. That is as important as what the AI should do. The AI should also have value. 

What to do if it doesn’t recognize the case? In a virtual world. The AI. It can make as many mistakes as the user allows. But in real life. When AI controls physical things. There is no room for errors. If the AI controls robot forklifts. Those systems can break lots of merchandise. If they work wrong. If the AI controls vehicles. like cars. And it reacts the wrong way. Results can be devastating. In real traffic, the vehicle has no time to wait and analyze opportunities. 

If we want to use virtual environments. The AI can wait. More information for the entire day. The virtual system. It can have endless time to try again. Or wait for more information. 

The world in the virtual environment. There, the system handles things like numbers. There are only two possible cases. Right (+) or wrong (-). But in cases like traffic, there are also plus-minus (±) cases. When AI controls a car. It can face a situation. That there is an emergency vehicle behind it. The AI can be ordered to drive to the sidewalk. The AI must also have orders that it must not impact people. And those cases. That don’t happen very often. They are the most challenging things for the AI. The AI must be prepared. That somebody tries to rob the car. Or there is an emergency vehicle behind it. In a tight avenue. 

Boolean algebra. 

The AI can learn in three main ways. 

1) Reinforcement learning

“In machine learning and optimal control, reinforcement learning (RL) is concerned with how an intelligent agent should take actions in a dynamic environment in order to maximize a reward signal. Reinforcement learning is one of the three basic machine learning paradigms, alongside supervised learning and unsupervised learning.” (Wikipedia, Reinforcement learning)

2) Supervised learning

“In machine learning, supervised learning (SL) is a paradigm in which an algorithm learns to map input data to a specific output based on example input-output pairs. This process involves training a statistical model on labeled data. Each input is paired with the correct output. The term "supervised" refers to the role of a teacher, or supervisor. Who provides. This training data guides the algorithm. Towards correct predictions. For instance, if you want a model to identify cats in images, supervised learning would involve feeding it many images of cats (inputs) that are explicitly labeled "cat" (outputs).” (Wikipedia, Supervised learning)

3) Unsupervised learning

“Unsupervised learning is a framework in machine learning where, in contrast to supervised learning, algorithms learn patterns exclusively from unlabeled data. Other frameworks. In the spectrum of supervision. Including weak- or semi-supervision, where a small portion of the data is tagged, and self-supervision. Some researchers consider self-supervised learning a form of unsupervised learning” (Wikipedia, Unsupervised learning)

“In supervised learning, the training data is labeled with the expected answers, while in unsupervised learning, the model identifies patterns or structures in unlabeled data.” (Wikipedia, Supervised learning)

“The typical framing of a reinforcement learning (RL) scenario: an agent takes actions in an environment, which is interpreted into a reward and a state representation, which are fed back to the agent.” (Wikipedia, Reinforcement learning) The thing that gives feedback. Like determining whether the case is favorable. Or non-favorable. It can be the human. 

The main problem with the AI and learning system is. How to determine whether the solution is good or bad. The simplest way is to use a human as a controller. When the algorithm ends its operation. Human operators. They select whether the solution is right or wrong. Determination of the desired solutions. It can also be programmed into the algorithm. In the case of stock marketing, desired. Or. A favorable solution could be maximized income. In the series of actions, the algorithm repeats the action. Time after time. The solution that it pursues. That is, maximizing income. 

Stock market analysis is a simple solution for modeling. The rising line, or rising income. It is the positive solution. The decreasing line is the negative thing. 

This type of machine learning is not hard to make. The user must only determine the highest number. That is, in a certain column. That is what the AI should pursue. In a series of cases, the user marks the wanted solutions, or actions. As positive (+) and negative (-). The thing. The algorithm must pursue. It is the highest possible number of positive solutions. 

The user determines the plus and the minus. And the AI tries to take as many points in the plus column. As possible. The AI, or its teacher, just selects the answer. That is marked as plus. The process requires more than one point. And then the AI follows the line. When the line is rising. The AI makes the right (+) solution. When the line decreases, the solution is wrong (-). 

https://vertexestechnology.com/levels-of-ai/

https://en.wikipedia.org/wiki/Boolean_algebra

https://en.wikipedia.org/wiki/Supervised_learning

https://en.wikipedia.org/wiki/Reinforcement_learning

https://en.wikipedia.org/wiki/Unsupervised_learning