Scientists have already proven that artificial intelligence can defeat a person in board games like chess or Go. Now it’s time to go further and swing at competitive video games. Today, the staff of the OpenAI research laboratory, founded by Elon Musk and Sam Altman, announced the creation of smart bots for Dota 2. Thanks to new technologies and accelerated learning mechanisms, a new generation of AI will be able to defeat the best entry-level players in battles 5×5. This is a big breakthrough because earlier this same laboratory was able to implement the bot only for a 1×1 fight. To try an improved version of the opponents are planning in August at the tournament The International with the strongest cyber sport of the world.
The creators of the intelligent bot explain their motivation by the desire to move further in the development of artificial intelligence. If you now learn how to win a machine in a team video game with a lot of variables, then you can scale the project’s capabilities and transfer them to events in real life. For example, you can similarly configure a computer to monitor the transport infrastructure.
This is an exciting event – we are moving to real applications! If you have a simulation of a task and you can scale it, then you will not have any problems along the way, “said Greg Brockman, technical director of OpenAI.
Video games were more difficult to learn AI because there are many more rules and variables. In chess, you can clearly identify the playing field, shapes and their capabilities, after which you only need to calculate all possible options. With Dota 2 is more difficult, because there is more information here. 115 heroes, each with four or more skills, items in inventory, mana and life indicators, level and characteristics. All this is tied to time and certain events on the map. And AI must be controlled not by one character, but by the whole team so that all these variables can be multiplied by five. It turns out more than a thousand different options for the development of events per second and more than 20 thousand parameters that affect what is happening.
Bots “trained” from scratch, then if no basic concepts and tactics were laid. The developers set goals and rewards for their achievement, and artificial intelligence had to work and learn. First, the characters just ran around the map, then they began to learn basic functions – killing creeps (monsters in Dota 2), buying items, exploring the territory and possible profitable positions. Naturally, the developers of the project accelerated the learning process. Every day the bots trained for 180 years of playing time in an accelerated mode. And it has been going on for several months.
Nevertheless, although the bots are learning at great speed, Dota 2 remains too complicated for them. The developers decided to lower the bar and selected only five heroes out of 115 (Necrophos, Sniper, Viper, Crystal Maiden and Lich), which AI can play. In addition, wards, invisibility were completely excluded, and the skills and objects are selected by the heroes according to a pre-planned plan. But the car has other trump cards in his sleeve – a lightning reaction, accurate hits, an instant check of the opponent and an assessment of the situation. Where a person thinks of a split second, a smart bot will act instantly. This summer, AI will battle with the strongest players of Dota 2. It will be interesting to see how professionals will fight with a specially trained computer.
I am a 24 years old boy looking forward to becoming the top blogger in the world. I work efficiently because my writing is influenced by the broader experience of the world, and I know how to work well with people and deadlines.