Centralized versus Decentralized Team Coordination Using Dynamic Scripting

Computer generated forces (CGFs) must display realistic behavior for tactical training simulations to yield an effective training experience. Tradionally, the behavior of CGFs is scripted. However, there are three drawbacks, viz. (1) scripting limits the adaptive behavior of CGFs, (2) creating scripts is difficult and (3) it requires scarce domain expertise. A promising machine learning technique is the dynamic scripting of CGF behavior. In simulating air combat scenarios, team behavior is important, both with and without communication. While dynamic scripting has been reported to be effective in creating behavior for single fighters, it has not often been used for team coordination. The dynamic scripting technique is sufficiently flexible to be used for different team coordination methods. In this paper, we report the first results on centralized coordination of dynamically scripted air combat teams, and compare these results to a decentralized approach from earlier work. We find that using the centralized approach leads to higher performance and more efficient learning, although creativity of the solutions seems bounded by the reduced complexity.