In this thesis, an action selection mechanism(ASM) is proposed for each agent given its role in multi-agent system for robot soccer and implemented in the real robot soccer game. Robot soccer game has very dynamic and uncertain characteristics because the ball and robots move fast and constantly. Furthermore, there exist competitions between agents of one team and hostile agents of the opposite team. It is a basic problem to determine what action the agent should take in the given situation. Particularly, the soccer-playing robot should take an appropriate action according to its role or position such as striker, sweeper and goal-keeper. Some computational mechanism for such action selection is essential in that sense. It is assumed that the role of each agent is given and the number of its available actions is finite. First, a relatively simple ASM is designed for the situation with no opponents. After that, the mechanism incorporates some additional action selection schemes to consider the opponents. The opponents are considered as a kind of disturbances for each agent. The ASM for each agent considers the opponents only when the level of disturbance is over prespecified degree. The proposed ASM consists of Internal Motive, Supervisor, Action Set, Intervention, and Final Selection modules. Particularly, Intervention module is implemented with neural networks. The efficiency and applicability of the proposed ASM is demonstrated under the real robot soccer game.