Evolving RoboCup2D Agents Based on PSO

摘要

In order to build intelligent robots to accomplish soccer game tasks, this paper introduces evolutionary computing in agent architecture for perception, planning, and action: (1) an architecture based on PSO is proposed, which made up of 4 levels: atomic action, combo action, behavior and policy. (2) by offline training, agents format perception rules and relevant parameters, to optimize perception method for the position, orientation and other information; (3) according to the granularity, functions, and parameters manually specified, PSO builds a set of combo actions, which described by atomic actions, parameters and execution results; (4) according to game environment and a few task rules, PSO searches for task, behavior, and combo actions, as a whole, to accomplish the game tasks. The simulation experiments on RoboCup2D platform show that, agent based on PSO is a robust and flexible robot control method: given evaluation methods and implementation frames, it is able to learn rapidly in real environment, and displays planning behavior without the use of classical planning techniques.