Using decision tree confidence factors for multi-agent control

Abstract

Ahhough decision t,rees are widely used for classification tasks, they are typically not used for agent control. This paper presents a novel technique for agent control in a complex multi-agent domain based on t.he confidence factors provided by the C4.5 decision tree algorithm. Using Robotic Soccer as an esample of such a domain, this paper incorporates a previously-trained decision tree into a full multi-agent behavior that. is capable of controlling agents throughout an entire game. Along with using decision trees for control, this behavior also makes use of the ability to reason *about action-execution time to eliminate opt,ions that would not have adequate time to be executed successfully. This multiagent behavior represents a bridge between low-level and high-level learning in the layered learning paradigm. The newly creat,ed behavior is tested empirically in game situations.