Artificial intelligence is getting smarter and gradually beats a person even in those games that require not only knowledge and cold calculation, but also intuition. So, in 2016, a computer beat a human for the first time in Go. In 2017, poker was under threat – specialists from Carnegie University – Mellon created a bot that challenged professional players.
Poker is a game that is very difficult to teach a computer to play: a good player quickly recognizes the strategies embedded in artificial intelligence and finds a way to defeat the bot. It is especially difficult for the computer if the bets at the poker table are unlimited, that is, the player can put an unlimited number of chips on his turn.
However, poker bots are a very popular direction for the development of the game. There are two types of poker bots. Some are quite simple and fight people in a low stakes game where the level of poker is very low and people can’t figure out even the simplest strategies. Such bots are not very interesting for science and serve to make money – poker sites, as a rule, try to fight them.
The second type is bots that compete with professionals. They are needed not only and not so much to make money, but to advance science. The topic of “games with incomplete information” is now one of the most popular in economics – it is no coincidence that Lloyd Shapley and Alvin Roth received the Nobel Prize in Economics in 2012 precisely for the theory of stable distribution, which is connected precisely with “game theory”. If a computer consistently learns to play games with incomplete information better than a human, we may no longer have to haggle and wonder if we made a mistake buying a new car with the characteristics we need at this price – because it will be decided for us by an app on a smartphone.
It is this motive that explains their interest in poker at Carnegie Mellon University. A year and a half ago, they created the Claudico bot, which took on professional poker players head-to-head with unlimited stakes. And lost.
Scientists did not despair and in two years they created an improved version of Claudico – Libratus. It is he who will fight with four poker players throughout January. At the same time, another group of scientists achieved a big victory over poker players – in 2015, a team from the University of Alberta (Canada) created a bot that perfectly plays one-on-one poker with limited stakes – even knowing the bot’s strategy, it was impossible to beat it: at best, a human brought the game to a draw or achieved an insignificant (at the level of error) advantage.
The conditions of the new tournament are as follows. Humanity in the tournament is represented by four different players who successfully play heads-up poker at the highest stakes: Jimmy “dougiedan678” Chow, Dong “DongerKim” Kim, Jason “PremiumWhey” Les and Daniel “ForTheSwarm” Mac’Olay. They must play against Libratus for 20 consecutive days of 1500 hands each. As a result, there will be 120 thousand such hands (each will play 30 thousand), which will allow you to collect statistics in order to conclude: who is cooler, a man or a machine?
Bookmakers believe that scientists are far from successful: bets on the victory of Libratus are accepted with a coefficient of 4 to 1 or even 5 to 1. However, the computer turned out to be better than people on the first day of the competition: Daniel McOlay and Jimmy Chow were in a small plus, and Jason Les and especially Dong Kim lost heavily.
How machine learning works?
How did artificial intelligence learn a complex card game? Just like solving many other problems – with the help of machine learning.
Everyone is talking about neural networks and deep learning, but neural networks are only a subfield of such a vast subject as machine learning. There are several hundred other algorithms that can solve AI problems quickly and efficiently, and in most cases are more human-interpretable. In this article, we will consider the algorithms of classical machine learning, the principle of operation of neural networks, the preparation of data for training models, and the tasks that are solved using artificial intelligence.
The main tasks of machine learning:
- regression recovery (forecasting) – building a model capable of predicting a numerical value based on a set of object features;
- classification – determination of the category of an object based on its characteristics;
- clustering – distribution of objects.
Let’s say we have a dataset with application statistics. It contains the following information: size, category, number of downloads, number of reviews, rating, age rating, genre, and price. Predicting an app’s rating based on features: size, category, age rating, genre, and price is a regression problem. Determining the category of an application based on a set of features: size, age rating, genre, and price is a classification task. Splitting applications into groups based on a variety of attributes (for example, the number of reviews, downloads, ratings) in such a way that applications within a group are more similar to each other than applications from different groups.
All this can be achieved by solving a large number of simple tasks with artificial intelligence. A convenient tool for such AI training is a platform created by Sypwai specialists. It is available to every user. Any of you can also take part in artificial intelligence training, earning money on it.
Registration on the Sypwai website
To start learning artificial intelligence using the Sypwai platform, you need to go to the official website of the company and register there. To register, fill out the proposed form, where you need to specify a phone number and email address, as well as enter login and password symbols. Double confirmation of registration via phone and e-mail is necessary to protect user data from leakage.