Modeling team performance for coordination configurations of large multi-agent teams using stochastic neural networks

Coordination of large numbers of agents to perform complex tasks in complex domains is a rapidly progressing area of research. Because of the high complexity of the problem, approximate and heuristic algorithms are typically used for key coordination tasks. Such algorithms usually require tuning algorithm parameters to yield the best performance under particular circumstances. Manually tuning parameters is sometimes difficult. In domains where characteristics of the environment can vary dramatically from scenario to scenario, it is desirable to have automated techniques for appropriately configuring the coordination. This research presents an approach to online reconfiguration of heuristic coordination algorithms. The approach uses an abstract simulation to produce a large data set to train a stochastic neural network that concisely models the complex, probabilistic relationship between configurations, environments and performance metrics. The final stochastic neural network, referred as the team performance model, is then used as the core of a tool that allows rapid online or offline configuration of coordination algorithms to particular scenarios and user preferences. The overall system allows rapid adaptation of coordination, leading to better performance in new scenarios. Results show that the team performance model captured key features of a very large configuration space and mostly captured the uncertainty in performance as well. The tool was shown to be often capable of reconfiguring the algorithms to meet user requests for increases or decreases in performance parameters. This work represents the first practical approach to quickly reconfiguring complex sets of algorithms for a specific application.