Most AI training follows a simple principle: match your training conditions with the real world. However, the new research of the Massachusetts Institute of Technology is challenging this basic assumption in AI development.
They discovery? When the AI system receives training in a simple environment, it usually performs better without predicting, rather than complex conditions facing in deployment. This discovery is not only surprising, but it is also likely to reshape how we think about building a more capable AI system.
The research team found this model while using classic games such as PAC-Man and PONG. When they train AI in predictable versions and then test them with unpredictable versions, it always surpasses AIS that is trained directly under unpredictable conditions.
In addition to these game solutions, the discovery has an impact on the future of AI development from robotics technology to complex decision -making systems.
Traditional method
So far, the standard and method of AI training follow clear logic: If you want AI to work under complex conditions, please train under the same conditions.
This leads to:
- The training environment aimed at matching real world complexity
- Test in a variety of challenges
- Under a large amount of investment under the training conditions for creating realistic reality
But there is a basic problem with this method: When you train the AI system with noisy and unpredictable conditions from the beginning, it is difficult for them to learn the core mode. The complexity of the environment interferes with their ability to master the basic principles.
This constitutes several key challenges:
- The training efficiency is significantly reduced
- The system is difficult to identify the basic mode
- There is usually no expectation
- Resource demand has increased sharply
The discovery of the research team proposed a better way to start with a simplified environment and make the concept of AI Systems Master Core before introducing complexity. This reflects effective teaching methods. In this method, basic skills create the foundation for handling more complicated situations.
Indoor training effect: violation of intuition discovery
Let us break down the actual discovery of researchers of the Massachusetts Institute of Technology.
The team has designed two types of AI agents for experiments:
- Acting agent: These have been trained and tested in the same noisy environment
- Summary: These are trained in a clean environment, and then tested in a noisy environment
To understand the learning methods of these agents, the team uses a framework called Markov Decision (MDP). The maps that can be regarded as all possible situations and actions that MDP can take by AI, as well as the possible results of these actions.
They then developed a technology called “injection noise” to carefully control these environments. This allows them to create different versions of the same environment at different levels of randomness.
What are the “noise” in these experiments? This is an element that cannot predict the result:
- Actions do not always have the same results
- How to move random changes
- Unexpected state change
When they tested, something unexpected happened. Summary drugs (people who are trained in a clean, predictable environment) are usually better handling noise than the agents of training specifically for these conditions.
This effect is surprising that the researchers named it “indoor training effect”, which is about how to train the traditional concept of the AI system.
The game is better understanding
The research team turns to classic games to prove its point of view. Why is the game? Because they provide the control environment, you can accurately measure the performance of AI.
In PAC-Man, they tested two different methods:
- Traditional method: Training AI with an unpredictable ghost action version
- New method: First train a simple version, and then test it in an unpredictable one
They conducted similar tests on table tennis to change the paddle’s response to control. What are the “noise” in these games? Example includes:
- Occasionally the ghost sent by the beans
- Pash on paddle that will not always react in table tennis
- How to move the game element how to move random changes
The result is obvious: AIS trained in a clean environment has learned more powerful strategies. In the face of unpredictable situations, their adaptability is better than the corresponding person who trains under noisy conditions.
These numbers are backup. Researchers found these two games:
- The average score is higher
- More consistent performance
- Better adapt to the new situation
The team measured things called “Exploration Mode”-how to try different strategies in the training process. AIS trained in a clean environment has developed a more systematic method of solving problems. It turns out that this is essential for future processing unpredictable situations.
Understand the science behind success
The mechanics behind the indoor training effect is very interesting. The key is not only about clean and noisy environment, but also how the AI system builds their understanding.
When agents explore in a clean environment, they will develop vital things: clear exploration mode. Think about it like building a psychological diagram. Without noise to cover up pictures, these agents will create better maps and no effective maps.
The study revealed the three core principles:
- Mode recognition: The agent in the clean environment recognizes the true mode faster, and will not be distracted by random changes
- Strategic formulation: They have established a stronger strategy, which will continue to be complicated.
- Exploration efficiency: They discovered more useful state actions during the training
The data shows the eye -catching things about the exploration mode. When researchers measured how the agent explores the environment, they discovered a clear correlation: agents with similar exploration mode performed better, no matter where they trained.
The impact of real world
The meaning of this strategy far exceeds the game environment.
Consider training robots for manufacturing: We can start with simplified task versions, rather than immediately throwing them into complex factories simulation. Studies have shown that they will actually better deal with the complexity of the real world.
The current application may include:
- Robot development
- Autonomous vehicle training
- Artificial Intelligence Decision System
- Game AI development
This principle can also improve the way we conduct AI training in each field. The company may::
- Reduce training resources
- Build more adaptive systems
- Create a more reliable AI solution
The next step in this field may explore:
- The best development from simple to complex environment
- New methods to measure and control environmental complexity
- Application in emerging AI fields
Bottom line
It was originally a surprising discovery in PAC-Man and PONG that it has evolved into the principle that can change the development of AI. The effect of indoor training shows that the establishment of a better AI system may be simpler than what we think-starting from the foundation, mastering the fundamentals, and then cope with complexity. If the company adopts this method, we can see a faster development cycle and more capable AI system in each industry.
For those who build and use the AI system, the information is clear: Sometimes, the best way is not to re -create all the complexity of the real world in the training. Instead, first pay attention to establishing a strong foundation in the controlled environment. Data show that powerful core skills are usually better adaptable in complex situations. Continue to observe this space-we just started to understand how the principle improves AI development.
