Designing intelligent agents that represent mobile robots in a complex environment is challenging, especially if the agents are to cooperate in a multiagent context and learn from experience. Depending on the specifications of the robots, the application requirements, and the type of environment, the design procedure will be different from one application to another.; No matter what the application is, the most important step in designing an intelligent agent is identifying its architecture, which shows the capabilities of the agent and reflects the nature of the environment.; Once the architecture of the agent is designed, modules and subsystems of the agent can be implemented according to the specifications.; In this thesis, the main application is robotic soccer in the category of small robots with an omni-scientific knowledge of the environment and without physical inter-robot communication. To make this application multiagent with all the necessities such as autonomy, concurrency and ability to communicate, an agent-oriented simulator was created to control small robots.; The agent architecture is designed to have two reactive and coordination layers to accommodate the rapid changes in the environment along with coordination with other agents. The reactive layer is equipped with a repository of behaviours that can make a direct connection from perception to action for a fast reaction to new situations. The coordination layer coordinates the local behaviours to choose the best behaviour according to the current state of environment and also coordinates the local decision of the agent with other agents to achieve cooperation and avoid conflicts.; Since, the agent is totally controlled by individual behaviours, its performance entirely depends on the behaviour design. In this thesis, a methodology is proposed and tested to facilitate the behaviour design procedure based on human experience.; Learning from experience, or in other words improving the quality of actions incrementally, is addressed by learning how to select behaviours in different situations. Although the individual behaviours are not adaptable, the arbitration among them is adaptable. The environment provides reward and/or punishment and the agent learns to choose those behaviours that bring reward, and avoid those that lead to punishment. Different reinforcement learning techniques are examined for this purpose.; As a necessary part of reinforcement learning and to deal with the continuous state variables in this application, a function approximation technique called Adaptive Fuzzy (AF) technique is introduced and compared with some other existing techniques.; All the agents with the hybrid architecture, equipped by properly designed behaviours, having adaptable arbitration unit, and able to communicate with each other were put together to play a full game. Simulations showed a cooperative social behaviour and proved that the agents were capable of learning from experience.
展开▼
