Classifying Adversarial Behaviors in a Dynamic Inaccessible Multi-Agent Environment

摘要

Developing intelligent agents for multi-agent, inaccessible, adversarial environments is arguably one of the most challenging areas in artificial intelligence today. Great strides have been made in developing emergent cooperation among teammates, but less progress has been made in quickly and automatically changing overall team strategy in response to adversary actions. One way that humans do such adaptation is by noting a similarity to a past adversary. This project is a system to do that sort of classification. The system is fully implemented in the simulated robotic soccer environment as used in RoboCup. The system does the following: Each agent observes the adversary and records relevant features. Based on these observations, each agent then classifies the adversary with regards to a set of pre-defined behavioral classes. The agents record their classification, and the team classification is decided by a simple majority. The effectiveness of this system on some simple behavior classes is shown. Future directions can include machine learning of behavior classes and strategy changes for those behavior classes, as well as developing more complicated classes.