Multiagent Systems is the subfield of AI that aims to provide both principles for construction of complex systems involving multiple agents and mechanisms for coordination of independent agents' behaviors. As of yet, there has been little work with Multiagent Systems that require real-time control in noisy, adversarial environments. Because of the inherent complexity of this type of Multiagent System, Machine Learning is an interesting and promising area to merge with Multiagent Systems. Machine learning has the potential to provide robust mechanisms that leverage upon experience to equip agents with a large spectrum of behaviors, ranging from effective individual performance in a team, to collaborative achievement of independently and jointly set high-level goals. Especially in domains that include independently designed agents with conflicting goals (adversaries), learning may allow agents to adapt to unforeseen behaviors on the parts of other agents.
Layered Learning is an approach to complex multiagent domains that involves incorporating low-level learned behaviors into higher-level behaviors . Using simulated Robotic Soccer (see Section 2) as an example of such a domain, a Neural Network (NN) was used to learn a low-level individual behavior (ball-interception), which was then incorporated into a basic collaborative behavior (passing). The collaborative behavior was learned via a Decision Tree (DT) .
This paper extends these basic learned behaviors into a full multiagent behavior that is capable of controlling agents throughout an entire game. This behavior makes control decisions based on the confidence factors associated with DT classifications--a novel approach. While operator success probabilities have previously been stored in tree structures , our work makes use of confidence measures originally derived from learning classification tasks. It also makes use of the ability to reason about action-execution time to eliminate options that would not have adequate time to be executed successfully. The newly created behavior is tested empirically in game situations.
The rest of the paper is organized as follows. Section 2 gives an overview of foundational work in the Robotic Soccer domain. The new behavior, along with explanations of how the DT is used for control and how the agents reason about action-execution time, is presented in Section 3. Extensive empirical results are reported in Section 4, and Section 5 concludes.