Alicja, Polish artificial intelligence allthingsblogging.com

Share
  •  
  •  
  •  
  •  
  •  
  •  
  •  
  •  

My fascination with artificial intelligence began with the book cycle "Felix, Net and Nick" by Rafał Kosik, where Manfred – SI appears regularly.

When in 2017 I discovered that in our city library there are books on AI, machine learning and its practical use, and Python textbooks, I started to learn and started the project.

From the beginning, I had a clear goal – to create a program that will deal with most common problems to speed up my work. Creating such a program required (and still requires) full commitment and persistence.

What is Alice?

Alice is a program capable of making decisions based on the collected data and learning by trial and error. The Ali core deals with only one thing: analyzing data packets from maintenance routines, i.e. checking that the task has been carried out in accordance with the guidelines and whether it is in accordance with the user's request.

Although it is not able to do anything in itself, except for analyzing data, it uses numerous helper programs to accomplish the task efficiently and keep the equipment in good condition. It is currently equipped with 48 such programs, which I will later call modules.

The first module I connected was the chess bot (16/03/2017), and by far the most interesting is the module responsible for creating graphics and plans, launched for the first time 27.

Okay, but how does it work?

At the Poznań Game Arena 2019 at the allthingsblogging.com booth, visitors got an extraordinary chance to challenge Alice. I taught her to play Carcassonne, a computer adaptation of a known board game. Armed with a powerful processor from Intel, it was a big challenge, which attracted crowds of fans of the original board game. It gave me a huge amount of information about the pace of learning and the direction of development of Ali. Ala began her adventure at PGA 2019 after playing 15,000 (!) Games with herself and 86 with a man (me). This knowledge gave her knowledge of the basic principles of the game and the simplest tactics. Since I won most games before PGA (83%), Alice adopted my habits because she found them profitable.

Fortunately, when the trade fair began and more and more players challenged, the percentage of wins began to change slowly.

On the first day 163 participants of the fair played with Alice, and my AI won 57 of these games, which gives a win percentage of 34%. Not a very impressive result, which, however, changed very quickly during the second day of PGA. This time, already armed with experience from games with many people using various tactics and representing different levels of skill and thinking styles, Ala played 343 games, of which she won 337. In this way, on Saturday, her effectiveness increased to an impressive 98%! On the third day, the statistics of the victories were also impressive: of 201 games played, Ala won 186 (92%).

The pace of teaching it can be demonstrated by the example of a multiplayer game that our crew on the second day forced. Ala was not prepared for her at all, without knowing any tactics. The first 4 games ended in a devastating defeat. My software scored only 0-4 points in them. However, in subsequent competitions, after analyzing the players' play, she was able to keep up with human opponents, scoring 10-20 points in each game, and even winning several times! Unfortunately, Sunday was the last day of the PGA, so the experiment had to be completed here.

Consumption of computing power

Unfortunately, such performance and rapid learning require considerable resources of computing power? How much? It depends on the task it has to perform and how much time it can devote to it.

Typically, on my laptop (with an 8-core Intel Core i7-4720HQ processor on board) with normal use of Alice (playing chess, arranging graphics / plans) consumes 60-70% of my processor's power. With more advanced activities (learning a human language, analyzing large amounts of data, installing on new hardware), the CPU usage stays close to 100%. In this context, it is interesting to compare Alice's consumption of computing power during her guest performances at the allthingsblogging.com booth at the PGA, where it was installed on a computer equipped with a 9th generation Intel computing unit.

The computer used by Alicja on PGA was equipped with an Intel Core i9 9900K processor. While playing Carcassonne with individual opponents at our stand, Alice used only about 12-18% of CPU power, which means that "you didn't even sweat". The performance of the new Intel processor really came in handy when we started experimenting with games in which Alice faced several opponents. It was a completely new situation for the algorithms, so they had to start learning very quickly, which was associated with intensive data processing. As I wrote above, Alice lost the first games to many players, but not because of the lack of computing power – while learning the CPU performance did not go below 70%, sometimes reaching almost 90%. An interesting indicator of how the processor's performance affects the performance of my software was the installation of Alice and its modules. This process also generated a 100% load on the 9th generation super-efficient CPU, but it lasted an order of magnitude shorter than on an older processor. | allthingsblogging.com

3

Comments