RoboCup Agent Learning from Observations with Hierarchical Multiple Decision Trees

It is a dif£cult task to hand-code optimal condition-action rules for software agents. A solution to this is reinforcement learning. In reinforcement learning, agents acquire the condition-action rules by learning from their experiences. However, acquisition of complicated rules might take a great amount of learning time and learning might not converge. To solve these drawbacks, an approach called learning from observations has been proposed in which learning in an agent is performed by observing human actions in the same environment of that of the agent. In our earlier work, we have applied this learning approach to the RoboCup software agent domain and adopted C4.5 as a learning engine. In this paper, we discuss a novel learning methodology that exploits the hierarchy structure of classes using hierarchical multiple decision trees, each generated by C4.5. Simulation results con£rm the superiority of the hierarchical multiple version of decision tress over a single decision tree, in terms of both generalization ability and agent's performance.