Sin Tong Jun, a researcher at the Faculty of Information Science and Technology, developed a tactic supporting system for small size robot soccer.
The robot soccer tactics simulation system covers formation design, strategic positioning of players, a situation-based intentional cooperative attacking tactic, an aid-obligatory array-based dynamic cooperative defending method and a genetic-reinforcement learning-based online tactics learning method.
In the formation design, you can set initial positions and movement ranges of each player in formations like 4-3-3, 4-4-2 and 4-2-4. For dynamic transformation of formations, you can select any formation table when necessary.
In the strategic positioning, the positioning of team players without a ball is modeled for 40-metre stonewall defense, which supports all-out attack and all-in defense and the numerical superiority of the team around the ball at the same time.
In the situation-based intentional cooperative attacking tactic, preconditions and stoppage conditions for passing a ball among players in such situations with high stochastic presence as a kick-off or a corner kick are designed so that intentional cooperative attacks among players can be made.
In aid-obligatory array-based dynamic cooperative defense tactic, an assistant player is selected when it is impossible for a single player to change the current situation so as to support unified cooperation among players, and problems of task sharing among players are solved to ensure dynamic defense.
The genetic-reinforcement learning-based online tactic learning is divided into tactic rule generation and tactic selection control. In tactic selection control, you control parameters for kicking, dribbling and passing. In tactic rule generation, determination of attacking and defensive situations, determination of the main robot and an assistant robot, and determination of ball handling of the main robot and the movement point of the assistant robot are completed by genetic-reinforcement learning so that tactic rules for a goal can be found automatically.
Formation designing and strategic positioning can increase shooting success rate by 12%. Intentional cooperative attacking tactic and dynamic cooperative defense tactic can improve the average attacking rate by more than 3% and the average ball possession rate by more than 10%. Genetic-reinforcement learning helps to find a new team tactic against opponents within 1000ms after the kick-off of the first half.
The system is available for small size Robot League and for control of intelligent robots working for automated factories. It also finds use in the measurement and assessment of techniques newly arising in the distributed AI field.
© 2021 KumChaek University of Technology