Agent-based Players for a First-person Entertainment-based Real-time Artificial Environment.
G. Michael Youngblood, Lawrence B. Holder · 2004
The necessity for improved players and opponents in first-person entertainment-based real-time artificial environments has inspired our research into artificial game players. We employed the approach of creating agent-based players, one using simple reflex-action based components and another based on a back propagation neural network, to interact in a modified Quake II environment. Our evaluation of these agents utilizes two metrics, a statistical measure and graph similarity based on clustering from player performance traces against interactive feature points in the environments—both comparing the agents to a sample of 20 human players. The reflex-based agent was able to complete 29 of 100 test levels with 73.1 % within statistical human-performance levels and 15.4 % similar to the human players, while the BPNN-based agent was unable to complete any levels. These results, including some additional findings that reveal the potential for a combinational approach, are inspiring an architectural vision of a multi-agent system to provide an advanced artificial intelligence engine for the embodiment of advanced artificial characters in gaming and simulation domains.