Training Methodology

In order to minimize the duration of the training, a maximum number of games to each team was established before the evolution took place. The teams were divided in groups and played amongst them. The best from each group were the “chosen ones”. This approach of dividing in groups was implemented to accelerate the evolutionary process.

To automate the system of evolution, all the teams were stored in a database with their history of games. In the end of each “tournament”, a script responsible for selecting the best teams was executed so the creation of the next generation could take place.

Tests And Results

Since the technique shown is somehow independent of language, we had to make the decision trees of the players become a compilable code in the language of the soccer game simulator. To achieve it, we developed a “translator” that wrote the content of the tree and its node in a compilable file. Since this article is to show/teach how genetic programming can be implemented we won’t get in details about the translator.

A small number of generations were enough for us to observe some interesting behaviors. In the second generation almost all the teams had some kind of a goalkeeper (what not necessarily would have happened, but it is right to think that it is an important characteristic of a good team), besides most of the best teams had a player that acted in the attack running with the ball.

In the figure below, we could see a game between a team from generation 0, which basically had a behavior where all the players went marking, which is known as “Kid Soccer” (this team was eliminated during the evolution), against a team from generation 2, which had a goalkeeper and a player near the opponent’s goal.

Because of the compatibility list implemented, the initial nodes chosen and the soccer simulator utilized, which was a simple soccer 2D with 5 players, we didn’t need to evolve the population of teams very far to achieve good results. Only 3 generations were enough for our teams to beat the simulator pre-built soccer teams.

Conclusion

In this implementation, Genetic Programming achieved its purpose. The “World” created started with many “incoherent” teams and as the evolution took place the teams were becoming better, because the players were acquiring different and interesting soccer behaviors.

Genetic Programming is a very suitable technique for soccer games, since in these games an optimal strategy has not yet been proved to be known. So, we can use a genetic programming approach to develop strategies for the teams.

As a result of the training, there will be a huge number of teams which can provide a great diversity to the game, also each generation can represent a level of difficulty and it would go as far as desired, since each new generation of teams is better than the previous one (the average of teams), generating a wide range of difficulty levels.

Also, if the nodes are properly developed, all the teams generated will have “human” behaviors, since the strategy is randomly developed by the computer.

Although this technique is suitable for soccer games, the quality of the teams would be greatly dependent of the quality of the terminal nodes and the range of the non-terminals.

If the non-terminal nodes don’t have enough range, the teams won’t be prepared for all the possible situations of the game, also if the terminal nodes are not well developed, the actions that the players have won’t be good. The quantity of nodes is also very important, since the speed of the evolution is directly linked with the number of nodes.