From cheating to machine learning--the general situation of soccer AI

From cheating to machine learning--the general situation of soccer AI

Author: ALEXJC

Translator: Rai Yonghao (Love flower Butterfly)

Original address: Http://aigamedev.com/questions/football-ai-cheating-machine-learning

This article is published in The Flower Butterfly Blog (http://blog.csdn.net/lanphaday), if reproduced, please retain the full text, including the retention of this statement.

When talking about the game of artificial intelligence, we usually pay more attention to FPS or real-time strategy games, in fact, the challenges of sports games and solutions are also worthy of attention. Although there is not much open learning available in the academic and gaming industries, we can still talk about this area of research in terms of the products that EA and Take2 have published each year.

This week's topic comes from Anthony, who is doing a postgraduate project related to football games:

"How do you see the AI of sports games?" I think the sport AI is more difficult because there is no such degree of freedom as developing a fantasy style game (because the player never has the chance to actually experience that kind of scene). [Translator Note: This section is a bit of a ghost painting God easiest, painting a dog to draw a cat the most difficult meaning. ]

Everyone has been involved in sports and they have a clear expectation of the game. Some people know where the relevant academic resources can let me go into the study. ”

In general, the AI of sports games is structurally similar to FPS or RTS games: The bottom is control and animation, the middle tier is individual decision-making, and the upper level is tactical. But from each level, it's a lot different from other games.

Screenshot 1:madden NFL 2008 has so far the strongest soccer AI

Role animation

The higher the player's expectations of a sports simulation game, the more important the player's movement becomes. The role of the game to move, hold the ball, and other roles of interaction, should be very realistic. These are all based on the character animation.

Looking at the research work done so far in this field, we can affirm the importance of animation in the game. Many animation projects are funded or funded by EA or other developers. If you read the paper from the SIG Graphics conference, you'll find that many of the technologies that use AI-related techniques are:

"Role Animation" in the article on the aigamedev.com and motion synthesis related articles.

Also refer to the papers published by Jovan Popovic of MIT and Zoran Popovic of the University of Washington.

There is still a lot of room for improvement in this area, but there is a lot of work to be done in motion synthesis. NaturalMotion is an animated middleware that devotes to soccer simulations (called backbreaker), and unfortunately it is far from mature enough.

Real moving paintings can only use motion capture in the present and foreseeable future. Instead of using a small piece of time to simulate a physical reaction, developers will instead use action capture to provide targeted behavior as quickly as possible.

Screenshot 2: The perfect cut-ball animation is technically very difficult

Individual behavior

In the middle layer of the sports game architecture, the individual AI is much simpler than the animation layer. It usually uses a combination of multiple manipulations to control roles, such as moving away from the opponent, advancing towards the target, and looking for open spaces.

The AI of this individual is very similar to the role AI of the traditional first-person or third-person shooter games. It often uses simple decision making techniques, such as a finite state machine, to perform the tactics assigned by the upper AI. There is not much to study in this field because it is easy to understand.

Reference reading:

Graig REYNLD's "steering behaviors" is the theoretical elaboration of its work, and Opensteer is the realization

Mat Bucklan's "Programming Game AI by Example" book contains a chapter on how to apply this technology to football games.

Screenshot 3: Player basic behavior implemented by finite state machine

Tactical AI

From a tactical point of view, sports games are much simpler than shooting games: There is only one course, not multiple hurdles, and no complex geometry. The only thing you need to consider is collaboration among the different members of the team. Although this is not a minor problem, it is possible to know from many games that AI developers need to do the most important thing: make the player behave well both in terms of offense and defense.

Now it seems that there have been a lot of research on the AI in sports games, especially the Madden NFL game. For years, developers have simply used the player's choice to cheat, and now use the machine learning algorithm to do the task.

This PDF document describes the AI history of the Madden game from the middle of the 90 's to the 2004.

You can also use the Alberta University's research program for statistical learning.

Screenshot 4: The initial position and the goal of each player are tactical elements

Have you ever looked into the AI of sports games? What do you think of the AI behavior of those games?