How game theory improves the accuracy and efficiency of artificial intelligence
Haebom
Game theory is a mathematical theory that studies how people and organizations make decisions in certain situations. This theory analyzes how each individual or group behaves to maximize their own interests, and how that behavior interacts with the choices of others. In other words, game theory studies how to find optimal decisions in situations that are influenced by the choices of other people. It may seem difficult to say this, but let's look at some examples.
In fact, it is also a theory explained by the prisoner's dilemma. In this situation, it is assumed that A and B must be in an independent environment where they cannot exchange information with each other.
Dolphins and whales work together to catch more fish. Each of these can choose to cooperate or not cooperate. If both dolphins and whales cooperate, both can get a lot of fish. However, if only one party cooperates and the other does not, the non-cooperative party gains more. If neither cooperates, both will get less fish. In this way, game theory explains how each choice affects the outcome.
If we look closer at workplace examples, colleagues compete with each other for opportunities for advancement at work. Here again, each person can choose to cooperate or compete. When everyone works together, the work environment improves and overall efficiency increases, allowing the company to give bonuses to everyone. On the other hand, if individuals choose to compete, personal profits may increase temporarily, but the workplace atmosphere may worsen in the long run.
If you explain it this way, you can also think about traffic jams by applying game theory. The traffic jam problem that arises when all drivers try to choose the fastest route can also be analyzed using game theory. If all drivers choose the shortest route, that route may actually be the slowest. On the other hand, overall traffic flow may improve if drivers choose to take a slightly longer route back.
The core of game theory is to understand how each choice affects each other and establish the best strategy based on this.
What does artificial intelligence have to do with game theory?
Game theory can be effectively utilized to improve the efficiency and accuracy of AI, especially large-scale language models. Traditional methods of training AI can sometimes produce inconsistent or biased results. However, by applying game theory, AI can make more accurate and fair decisions by considering different scenarios and possibilities.
For example, it is designed to increase consistency between two systems within a language model: the 'generator', which generates questions, and the 'discriminator', which evaluates the generated answers. For example, for the question " What is the capital of France? ", the generator initially determines that there is an 80% chance of answering 'Paris'. The generator then flips a coin to decide whether to answer truthfully or falsely, and this choice affects how the discriminator evaluates the generator's answer.
The discriminator determines whether the answer provided by the generator is true or false, and if judged to be true, scores are given to both sides. On the other hand, if it is judged to be false, no points are given. This process is repeated approximately 1,000 times, and with each iteration, the generator and discriminator learn and adjust each other's responses. Through this interaction, the two systems gradually come up with consistent answers, which greatly improves the consistency and accuracy of the overall model.
This game helps language models consistently provide the same answers across different question formats. Nash equilibrium was chosen as a way to increase the reliability of the language model and make the model's answers more trustworthy to users.
Nash Equilibrium: This means a state in which both players cannot achieve better results by changing their strategies.
Researchers at the Massachusetts Institute of Technology (MIT) have developed a method called the ‘consensus game’. This is a game that forces the model to find an answer it can agree on in two modes: generative and discriminative. This game allows the model to improve its accuracy and internal consistency.
Personally, I found this paper while reading the ICLR paper, and it was a paper that won an award at Newlips last year. I feel this more and more these days, but as artificial intelligence operates 'like' humans, it seems like a good example of how applying things used in social science can yield good results.
