A growing practice in artificial intelligence training is the use of video games. Both Dota 2 and StarCraft have proven to be very popular platforms, and there was even a case where they used Mortal Kombat. Today we encounter Google Research Football, an open-source environment created by the Google Brain team to apply reinforcement learning to virtual agents with a single objective: Play football.
Although most general seasons have already ended or are on pause, the "football world" continues to be full of news. Portugal added another international title, there are two world cups in progress (women's and U-20), and the Copa America starts next Friday with two peculiar guests (Qatar and Japan).
However, today we are more interested in football crossing paths with artificial intelligences, and quickly the question arises: Can a virtual agent learn the details and complexities of football? A special team from Google Brain decided to find out, and the result is the environment Google Research Football, based on reinforcement learning. Here it is in action:
Google Research Football: A 'FIFA' for AI
At the heart of Google Research Football we find an advanced simulator called Football Engine, which is a modified version of the Gameplay Football project. The virtual agents participating in each match have the ability to control one or all players at once, and are subjected to normal football rules, with corner kicks, fouls, cards, offside and even penalties. The same extends to human players, thanks to support for keyboards and gamepads (the official page warns that actions have a 100 ms lag).
The Football Engine has high levels of optimization in C++, which translates to relatively low hardware requirements. For example, a computer with a six-core processor can deliver a performance of 25 million daily steps, without a GPU. The algorithms used were DQN and Impala, under two reward parameters: 'Goals scored', and 'goals plus proximity of the ball to the opposing goal'.
The main opponent for both algorithms was a hand-designed model with three difficulty levels: Easy, Medium and Hard. A DQN agent with 20 million steps managed to defeat the Easy model, while the Hard level could only be overcome by Impala (a distributed algorithm), using a minimum of 200 million steps.
Do you want to start training the next virtual Mohamed Salah? Then, download and compile the code on a system with Linux and Python3. The official profile of Google Research Football has instructions for installing dependencies and running experiments. However, those responsible left no doubt: This is not an official Google product, and we should not expect a football game with super AI in the future.
Official announcement: Click here
https://old.neoteo.com/problema-de-la-television-con-ruido-inteligencia-artificial/
